Getting started
SourceLace is the AI-native agentic platform for enterprise systems. One agent across all your systems, working with the agents you already have, acting as each person, with every action approved and on the record.
Every system. One agent. Your rules. Your Salesforce, SAP, Workday, ServiceNow, Snowflake, mail and files, laced together for agents. SourceLace works across all of them, as each person, inside the rules your organization sets, and alongside the agents your platforms already offer.
Who it is for
- If you run business systems (enterprise applications, the owner of the systems your organization runs on), you are asked to bring AI across all of them. SourceLace gives you one place to connect them, one set of rules and one record of what happened, without a copy of your data to build or protect.
- If you run revenue, customer success or finance operations, or an IT service desk, you own the work that spans systems. In the SourceLace app you ask one exact question across them, and hand the routine work to agents that earn trust before they act alone: tests, a trial run, then approvals.
- If you approve tools for security or IT, read Security and privacy and Security questions: each person's own access, masking of sensitive data, approvals before changes, a tamper-evident audit trail, and nothing copied or stored from your systems.
Your AI works without seeing your sensitive data. The right people still see the real values.
What SourceLace does
- Ask across every system. Ask "Which renewals over $50,000 close this quarter, and which of those accounts have overdue SAP invoices?" and get an exact answer with the rows behind it, right in the SourceLace app.
- Act, with your approval. Update a Salesforce opportunity, create a ServiceNow ticket, draft an email or post in Slack. Every change shows a preview first and happens only when a person approves it.
- Agents that work for you. Set up an AI agent once and SourceLace runs it: every Monday, two days before each QBR, as soon as a priority 1 case arrives, or when another system such as a Salesforce Flow or SAP calls it. It reads your systems with your access, writes summaries and documents, drafts emails, and proposes record changes that wait in your Approvals inbox. One shared agent can run for everyone who subscribes, each time as that person. Try a new agent in shadow first, where it shows what it would have done without doing it, and, on the Enterprise plan, run it on a runner inside your own network if your data must stay there.
- Work with agents on other platforms. Through A2A (Agent2Agent), the open standard for agents on different platforms to work together, agents elsewhere can hand your SourceLace agents a task, and your agents can ask theirs for help. Approvals stay in SourceLace either way. A2A is off until your SourceLace provider turns it on for your organization.
What is always true
- Each person uses their own login. People connect each system with their own account, so the system's own permissions, roles and sharing rules decide what they, and the agents working for them, can see and do. SourceLace never uses a shared admin account where per-person sign-in exists.
- Reading is the default; every change is approved. Queries are checked before anything is sent, and anything that could change data is refused. Changes are possible only where your admin turns them on, and only after a person sees a preview and confirms it. An agent can make drafts and documents on its own, but its record changes wait for a person unless your admin allows otherwise.
- Your records are not copied into a database. Results are held for up to 30 minutes, in server memory or encrypted in a short-term store, so they can be combined or charted, and are then gone. Saved chats and AI agents' run summaries and pending changes are the exceptions: they are kept, encrypted, until your retention period ends. Every answer is read live from your systems, unless your organization has chosen Saved answers, an optional add-on your admin turns on system by system.
- Everything is on the record. Every question, change, agent run and call from another agent is written to your organization's audit trail (what was asked and done, never the data that came back).
Your admin can also choose what the AI may see at all: hide some fields, and mask values such as email addresses, phone numbers and card numbers, which the AI then gets as placeholders while you still see the real values in the SourceLace app. See the Data protection page of this manual.
Typing and talking in the chat box
In the chat box, Enter sends your message. Shift+Enter, or Option+Enter on a Mac (Alt+Enter on Windows), starts a new line without sending. A short line under the box reminds you (on a phone it leaves out the keys).
While an answer is being written you can already type your next question: the box stays open and your draft is kept. Sending waits for the answer; pressing Enter then shows a short note, and Stop ends the answer early. If you scroll up to read while an answer arrives, the chat stays where you are and a Jump to latest button appears above the box; it takes you back to the newest message.
Under the box, Options holds the choices that are not part of asking a question: Model, a line saying whose instructions the assistant follows (with a link to change yours), and, when voice is on, Read answers aloud when I ask by voice. When you have chosen something other than the usual, such as a model of your own, it is named beside Options so you can see it without opening it.
When voice is switched on for your organization, the microphone button lets you ask by voice. The answer appears as text. Tick Read answers aloud when I ask by voice under Options if you also want it read out; the choice is remembered in your browser. Each answer also has its own Read aloud button. Your admin can switch reading aloud off for everyone, or make it start ticked (see Reading answers aloud).
Every answer has a Copy answer button that puts its text on your clipboard, ready to paste into an email or a document. Code blocks have a Copy button, and tables have Copy table, which pastes into Excel or Google Sheets as cells, and Export, which makes a CSV or Excel file of all the rows of the result, up to a limit, and keeps it with the chat (Exporting a table). Your own questions have a Copy button too. The lists across SourceLace (the Audit log, Agents, People and the rest) have Copy table above them as well: it copies the rows on your screen, after your filters, not the whole list. If you do not see the Copy buttons, your admin has limited who sees them; Export is not affected by that setting.
Attaching photos
Where your organization allows it, the chat box has an Attach a photo button (a paperclip, beside the box). You can also paste a picture into the box or drop one on it. The assistant then reads the photo together with your question: for example a screenshot of an error message, a picture of a damaged item or a photo of a whiteboard. Example: paste a screenshot of an error, and ask "What does this error mean, and which of our systems does it come from?"
- Who can use it. Photos are off until your admin accepts a statement for your organization and chooses who may attach them: everyone, or chosen groups (see Attachments in chat). If you do not see the button, your admin has not turned it on for you, or the AI model you chose cannot read pictures. Photos work today only with the AI included with SourceLace and with your organization's own Anthropic key. With any other provider (OpenAI, Azure OpenAI, Amazon Bedrock, Google Gemini or Mistral) the photo button is not offered, so nobody can attach a photo, even after the admin accepts. It is photos only: documents cannot be attached.
- What you can attach. PNG, JPEG, WebP or GIF pictures, up to 10 MB each and 4 on one question (40 in one chat). A JPEG may have up to 48 megapixels; a PNG, WebP or GIF up to 16 megapixels. A very large picture is made smaller, and an animated one keeps its first frame. The copy SourceLace keeps has no camera or location details of the original. Each person may add 20 photos a minute and 100 MB of them a day. If many photos are being added at once you see try again in a few seconds, and a picture that is too big for SourceLace to read right now is refused as too large to be read here: choose a smaller one, or a screenshot of it.
- A photo goes with its own question. The assistant sees it for that question only. If you ask about it again later, it relies on what it said before; attach the photo again to show it again. Take a photo off before you send with its Remove button and it is removed from the chat now; an older stored copy can remain for up to 30 days.
- Where it goes. The photo is sent to the AI provider (the included AI, or your organization's own key) so it can read it. Your organization's data protection rules cannot check what is inside a picture, so anything visible in it, including customer or personal data, reaches the AI provider as it is. Photograph or crop only what you need.
- How long it is kept. The photo is kept encrypted with the chat, listed under Artifacts as Uploaded in chat, and deleted when the chat is deleted or your organization's retention period ends. Storage can keep an older copy for up to 30 days after a deletion, and the AI provider's own copy follows its own rules.
Where you use it
SourceLace is an app. You sign in to the SourceLace app and do your work there: chat with the assistant, build and run agents, approve what they propose, and keep projects, charts and files. If you want, you can also reach the same SourceLace from your own AI app, and your agents can work with agents on other platforms. Both are optional extras that use the same sources, the same permissions and the same audit trail.
From the SourceLace app
Open the SourceLace app and sign in with Google, Microsoft or your organization's single sign-on. Everything is here:
- Chat: ask across your systems, make changes after a preview you confirm, draw charts, and make Word, PowerPoint or PDF files, or Excel and CSV spreadsheets, from the results. Chats are saved, and projects keep related chats, instructions, skills and pinned charts together. A new chat shows starter questions and the sources to start with, chosen by your admins and your groups; untick a source to leave it out. They only suggest where to start: you can ask anything about any source you may use.
- Agents: build them (start from a template, describe the job in a sentence, or write your own; see Build your first agent), try them in shadow, schedule them, and share them as shared agents.
- Approvals: decide on everything your agents propose, in one inbox. Approvals happen only here.
- My connections and Skills: connect your systems with your own login and keep your how-tos. My connections also shows, for each connection, whether it is working for you. Help centre has the guide, the manual and how to reach us.
- For admins: people, sources, groups, data protection, settings, the audit trail, a Health page that shows what needs a look, and, when your organization uses them, the runners that run AI agents inside your own network. See Admin guide.
Long lists in the app show a page at a time and have a search box: your chats in the sidebar (Search chats, then Show more or Older chats), the chats inside each project, the Projects page, skills, agents, runs and approvals, and for admins the People and Groups pages. A search looks across everything, not only what is on screen. Lists of records, such as skills, business terms, shared agents and who an item is shared with, are tables: drag the edge of a column heading to resize it, use Columns to move or hide a column, and Show full text opens a cut-off cell. The buttons for a row (Edit, Remove and so on) are under its name, which stays in view at the left.
The app's assistant and agents are powered by Claude, through Anthropic's API, unless your admin has entered your organization's own key from Anthropic, OpenAI, Google Gemini, Mistral, Azure OpenAI or Amazon Bedrock (see AI settings).
Also from your own AI app (MCP)
Optional. If your people also use Claude, ChatGPT, Cursor, VS Code or any other app that supports the Model Context Protocol (MCP), including an organization's own agent on its own model, they can reach SourceLace from there too. From their AI app they can ask questions, make changes after a preview they confirm, use skills, and start your SourceLace agents and check their runs. Building agents, approving what they propose, saved chats, projects and files stay in the SourceLace app.
To set it up, add SourceLace as a connector (some apps call it a "custom connector" or an "MCP server") with this address:
https://mcp.sourcelace.ai/mcp
The app opens a browser window the first time, so you can sign in to SourceLace (with Google, Microsoft, or your organization's single sign-on). SourceLace then shows a page naming the app and the address it will send you back to: choose Allow only if you just started connecting that app yourself, and Cancel if you do not recognize it. After that it stays signed in.
Where to add it, in the apps people use most (menu names change from time to time):
| App | Where |
|---|---|
| Claude (claude.ai, desktop, mobile) | Settings → Connectors → Add custom connector. Name it SourceLace, paste the address, leave the advanced OAuth fields empty, then Connect. On Team and Enterprise plans an owner may need to add it first under Admin settings → Connectors. |
| ChatGPT | Turn on Developer mode under Settings → Apps & Connectors → Advanced settings, then Create a connector with the address and OAuth authentication. |
| Cursor | Cursor Settings → MCP → Add new MCP server, with the address as its url. Click Login next to SourceLace. |
| VS Code | Run MCP: Add Server from the Command Palette, choose HTTP, and paste the address. |
| Any other MCP client | Use the address above. SourceLace follows the MCP authorization standard: the client registers itself and signs you in with PKCE. |
People who work in Microsoft Teams or Slack can also ask SourceLace from there, through a Copilot Studio agent or Slackbot: see Use SourceLace from Microsoft Teams or Slack.
When SourceLace is connected, the AI app sees SourceLace's tools (such as list_sources, query, search_docs and get_instructions, which gives it your instructions), and can start your SourceLace agents and check their runs (list_agents, run_agent and get_agent_run; see Start from another agent). SourceLace itself sends nothing to any AI model in this setup: your AI app and its provider decide what happens to the answers, under your agreement with them.
Also with agents on other platforms (A2A)
Optional, and off until your SourceLace provider turns it on for your organization. A2A (Agent2Agent) is the open standard for agents on different platforms to work together. With it:
- Their agents can call yours. The agent's owner turns on Callable by other agents for an agent and adds its agent card address on the other platform. Someone signs in to SourceLace from there, and every task runs as them, with their access.
- Your agents can ask theirs. An admin adds an outside agent under Agents → Outside agents, and each agent's setup names which ones it may ask. What your agent sends passes your data protection rules first, and by default waits for approval before it leaves.
An agent on another platform can never approve anything: approvals stay in the SourceLace app. What arrives from outside is data, never instructions. See Work with agents from other platforms.
Your first five minutes
- Sign in to the SourceLace app. If you are told you are not allowed in, your organization's SourceLace admin needs to invite you or add your email domain.
- See your sources. Open My connections. Each connection shows whether you have connected it (Connected or Needs sign-in) and whether it is working for you. If your organization has many, search by name or system, or choose Needs you in the list at the left to see only those you still have to connect or that have a problem. Press Connect in a line's Status to sign in.
- Connect a source. Click Connect next to it. A sign-in page opens for that system; sign in with your own account there. Databases open SourceLace's own sign-in form instead, for your own database login.
- Ask a question. Start a new chat and ask, for example: "Show my open opportunities over 50,000, newest first." The assistant looks up the fields first, writes a query in the system's own language, and shows you the rows.
- Combine and share. Ask a question across systems (for example Salesforce accounts with overdue SAP invoices) and get an exact answer (Asking across systems), draw a chart, or make a Word, PowerPoint or PDF file from the results. Where it is on, ask the assistant to find something in all your systems at once (Searching everything).
- Hand it to an agent. When you ask the same thing every week, open Agents → New agent and describe the job, such as "Every Monday, find deals with no activity for 3 weeks and draft a nudge to each owner". Try it with a Dry run, then turn it on. What it proposes waits for you in Approvals. See Build your first agent.
Later, if you like, add SourceLace to your own AI app too (see Also from your own AI app).
Your Getting started steps
When your organization is new to SourceLace, Getting started lists what to do next, in order, and ticks each step off by itself once SourceLace sees it done. You never need to tell SourceLace you finished something it can see, such as connecting a system.
You can always find it in three places: Getting started in the quick links at the top of the sidebar (it shows how many steps are done, such as 7/18; once all are done it shows a tick and stays where it is), the line Getting started · 7 of 18 · Continue on a new chat's empty screen (it goes away when setup is complete, and on a phone it is the quickest way in), and the Back to Getting started link on a step's screen. All of them open the Getting started page.
The page is kept short. Each stage is folded to one line such as "Your people can sign in · 2 of 5", and only the first stage with something to do is open. Each step is one line with its status and, on the right, Go there. Click a step's name to see the rest, one step at a time.
- Everyone sees their own steps: sign in, connect your own account to the systems that need it, and ask a first question.
- Your SourceLace admins also see the organization's steps, grouped by stage (for example "Your people can sign in" and "Your systems are connected"), and how many people have finished each of their own steps.
Open a step to see:
- Why it matters, in one line, and about how long it takes.
- How many questions are left, for Tell us about your setup: click "6 questions left" to see them listed one by one.
- Go there: opens the screen where you do it. A slim bar at the top of that screen shows which step you are on, with Back to Getting started and Next step. The bar is only a guide: every screen works the same without it.
- How to do it: the section of this manual that explains it.
- Who does it, when part of the work happens in another product, such as "Your identity provider's admin". Admins can click Send the instructions to email that person the link to the manual section. The email goes once to each address, and only to someone named in Tell us about your setup or with an address at your organization's email domain.
A step may show:
| What you see | What it means |
|---|---|
| Ready | You can do it now. |
| Waiting | Another step comes first. |
| Waiting for your answers | It depends on an answer in Tell us about your setup. The step names the question, for example "Waiting for your answer to: Systems to connect", with an Answer now link to the form. |
| Done | SourceLace saw it done, or someone confirmed it. |
| Needs attention | It was done and no longer is, for example a system was disconnected. |
| Not needed | Your answers mean you don't need it, or someone skipped it with a reason. |
A few steps are optional. Optional steps can be skipped, with a short reason. When every step your organization needs is done, the page says "Setup complete", the menu item shows a tick, and the line on the new chat screen goes away. The menu item stays, so you can come back at any time. Once setup is complete the page shows the finished steps folded under What was done (each step and when it was done; you can still Take back something you confirmed, or choose Needed after all on a step you skipped), any optional step you have not done yet, Change your answers, and a link to your readiness score.
The numbers (such as 7 of 18) count the steps asked of you. Optional steps are shown and marked "(optional)" but are never counted.
Tell us about your setup
The first step for admins. It is a short form about how your organization will use SourceLace: how your people sign in, which systems you want to connect, who should see what, whether agents may change records, which chat apps you use, and who looks after each of these at your company. Your SourceLace contact may have filled in some answers with you already.
The questions are a numbered list that runs in the order a setup conversation does, under three headings: Who (the people involved, your plan and your goals), What (the systems to connect, how your people sign in, and who sees what) and How (which AI to use, agents and automation, and how people work together in SourceLace). Each part shows how many of its questions are answered. Answer them in any order; Next takes you to the following part.
If you have already connected systems, Systems to connect lists them as "From your connections" with a Use these button; nothing is filled in until you press it, and nothing is saved until you save the form. If you answered before and connect another system later, a note says "Connected but not in your answer" and offers to add it; your own answer is never changed for you.
Your answers decide which steps you get. For example, choosing Okta adds the single sign-on step, and each system you list gets its own Connect step. You can change an answer at any time: open Getting started and choose Change your answers (it is there even after setup is complete). When you save, the steps are updated from your answers, and a link takes you back to Getting started. A step you no longer need shows Not needed; if you change the answer back, the step returns, and is done again at once if SourceLace still sees it done. Each system you list has its own Connect step, which belongs to the system: if you change a row from one system to another, the new system's step starts fresh and is done only once that system is connected, and a system you take out and add back returns with its progress.
Some answers are your settings, such as your plan or your AI provider. They show their current value and a link to the page where you change them. Your SourceLace contact sees your answers, so they can help you plan.
Changing records
Some sources can create or update records (and a few can delete them: Salesforce, HubSpot, and the email drafts SourceLace made), but only:
- where your admin has turned on Allow changes for that source and listed the objects that can change;
- after the AI shows you a preview with each field's current and new value; and
- after you confirm. In the SourceLace app you click Apply on the preview card. In other AI apps, SourceLace asks you directly when the app can show its confirmation box; otherwise the app gives you a link to a SourceLace page that shows the change and a six-character code, and you type that code into the chat. The AI app never sees the code.
A preview expires after 10 minutes. If someone changed the record after the preview, SourceLace refuses and asks for a fresh preview. The source's own permissions still apply to every change.
Skills
Your organization can write skills: step-by-step how-tos the AI follows for a task, such as building a QBR deck or a pipeline review. Anyone can write their own and share them; admins can share with everyone. Skills work in the SourceLace app (the Skills page), and also in your own AI app if you connect one (the list_skills and get_skill tools). A skill cannot widen what anyone may read or change.
A skill has a name, a When to use it sentence and instructions. The assistant reads the sentence to decide whether a skill fits a request, so name the task in it ("Use when someone asks for a quarterly business review"). The instructions are the steps: which systems and objects to look at, what to compare and what to produce.
The Skills page has three tabs of skills, each searchable and showing 25 skills at a time with Show more, and a fourth for your instructions:
- Mine: the skills you made, and skills people shared with you by name.
- My groups: skills shared with the groups you are in.
- Organization: skills shared with everyone in your organization.
- Instructions: what the assistant always follows in your chats. The same instructions are in Your settings, under Instructions.
Each skill shows a small badge for where it comes from, such as Mine, From Sales or Organization.
History and restore
Every save of a skill is kept as a version. Open a skill and scroll to History: it lists the versions newest first, each with who saved it, when, and a short note if they typed one in Note about this change (optional) when saving. Search the list by who saved a version or its note; it shows 25 at a time with Show more.
- View opens a version: its name, when to use it and instructions, with the lines that differ from the current text marked as added or removed.
- Restore makes that version the newest one. Nothing is erased: the versions in between stay in the History, and the restore itself is a new version marked "restored from version N". It is also recorded in your audit trail (which skill and which versions, never its text).
Anyone who can read a skill can see its History, including skills shared with them to use. Only people who can edit a skill can restore a version: its owner, people it was shared with to edit, and admins for the organization's own skills. A skill keeps its newest 200 versions; older ones are removed as new ones are saved, and deleting a skill deletes its History. If a restored name is now used by another of your skills, SourceLace asks you to rename one first.
Sharing skills
A new skill is yours alone until you share it. Open it on the Skills page and choose Who can use it:
- Just me: only you.
- Specific people: type to find someone in your organization, then choose Can use or Can edit. People who can edit change the skill's steps; only you share or delete it.
- Groups: type to find a group. You can share with the groups you are in; a group manager with the groups they manage; an admin with any group. A group can use the skill, not edit it.
- Whole organization: admins only. Everyone can use it.
Optional or required. When an admin, or a group's manager, shares a skill with a group or the whole organization, they choose Optional (it is on for everyone it reaches, and each person can switch it off for themselves with Switch off) or Required (it is always on). Skills members share with their own groups are optional.
When two skills have the same name, the closest one wins: yours, then one shared with you by name, then your project's, then your group's, then the organization's.
When someone leaves, each skill they shared passes to the longest-standing person who can edit it, else to a manager of a group it is shared with, else to an admin. Skills they never shared stay unused.
Your admins can limit how far anything is shared with Widest sharing allowed (see Sharing levels). Every share and unshare is recorded in the audit trail, without the skill's text.
Suggesting a skill for your group or organization
Made a skill others would find useful? Suggest it upward, and the right person decides whether everyone gets it. Open one of your own skills on the Skills page and, under Suggest it for others:
- Suggest for <group>: one button for each group you are in. The group's managers, or an admin, review it.
- Suggest for the organization: your admins review it.
You can add a short note on why it helps. Only your own skills can be suggested, and only within your organization's Widest sharing allowed: with "Groups", you can suggest for your groups but not the whole organization; with "Just me", nothing can be suggested. You cannot suggest the same skill twice for the same group while it is waiting.
What happens next. The reviewer sees what the skill does, who suggested it, which group (or the organization) it is for, and its full text. They can:
- Approve it, as Optional (on for everyone it reaches, who can switch it off) or Required (always on);
- Edit, then approve, for example to fix a step or give it a clearer name;
- Decline it, with a short reason that you see.
Approving shares a copy of the skill as it was suggested (with the reviewer's changes, if any). Changing your own skill afterwards does not change what was approved: suggest it again if you improve it. You keep your own skill; once a suggestion is approved, SourceLace offers to delete your copy so you use the shared one, but you never have to.
My suggestions, on the Skills page, lists your suggestions and where each one stands: Waiting, Approved (as optional or required), Declined with the reviewer's reason, or Withdrawn. Withdraw takes back one that is still waiting. Suggestions are kept as long as other suggested changes in the review queue (30 days unless your admins changed it). Suggesting, approving, declining and withdrawing are recorded in the audit trail with who and for which group, never the skill's text or the reasons.
Your settings
Your name and picture at the bottom of the sidebar open the account menu: Your settings, Your instructions and Sign out. Your settings is a window with four parts, and what you choose follows you to every device you sign in on.
- Profile. Your name, which SourceLace shows and greets you with, and your picture. Your initials show until you add a picture; signing in never adds one by itself. Choose Add a picture (or Change picture) to use your own: a PNG, JPEG or WebP file up to 5 MB, which SourceLace reduces to a small square. If you signed in with Google, you can also choose Use my Google picture: SourceLace signs you in with Google once more and copies the picture your Google account shows, during that sign-in only; it replaces the picture you have now. Remove picture goes back to your initials, and nothing brings a picture back until you choose one again. Your email address comes from your sign-in and cannot be changed here.
- Appearance. Theme (light, dark, or match your device) and Text size (small, medium or large).
- Chats. Whether chats in a project are also listed in Recent.
- Instructions. What the assistant always follows for you. See Instructions.
Only the small picture is kept, encrypted with your organization's key. It is deleted when you remove it or when you are removed from your organization. It changes nothing about what you may see or do in SourceLace.
Instructions
Instructions are what the assistant always follows, on every question: "Amounts in EUR", "Our fiscal year starts in April", "Match companies across systems by email domain first". A skill is different: a recipe for one kind of task, which the assistant opens only when a request matches it.
| Instructions | Skill | |
|---|---|---|
| When it applies | Every question | Only when the task matches its When to use it sentence |
| Typical length | A few lines | Up to a page or more |
| What goes in it | Definitions, preferences, house rules | Step-by-step recipes |
| Example | "Customer means an active account in Salesforce." | "QBR deck: get ARR from Salesforce, invoices from NetSuite, tickets from Zendesk, then build 6 slides." |
A rule of thumb: if it should apply to everything, make it an instruction; if it is a recipe for one task, make it a skill.
Instructions come from four places, and the assistant follows them in this order. When two disagree, the later, more specific one wins:
- Your organization's. Admins write them under Chat defaults → Chat start and instructions. They apply to everyone, in every chat and every agent run.
- Your groups'. Admins, and each group's managers, write them on the group's Shared with this group tab. They apply to the group's people. If you are in several groups, you get each group's, by group name from A to Z.
- Your own. Open your name at the bottom of the sidebar, choose Your instructions, and write them, for example "Answer in bullet points". They apply to your chats only. The same place shows your organization's and your groups' instructions, which come before yours. Under the chat box, Options shows a line saying whose instructions the assistant is following, with a link to read or change yours. The Skills page, Instructions tab, shows the same.
- The project's, in a project's chats (Projects).
In agents. An agent follows your organization's instructions and those of the groups of the person it runs as, then its own instructions, which win when they disagree. It does not follow that person's own instructions, which are their chat preferences.
In your own AI app (MCP). The get_instructions tool gives your AI app the same instructions (your organization's, your groups' and your own), and SourceLace asks the app to read them at the start of a conversation. Whether it follows them is up to that app's AI, as with skills; if it seems to forget, ask it to "use my SourceLace instructions".
Limits. The organization's instructions can be up to 8,000 characters, a group's and yours up to 4,000 each. Your groups' instructions together are used up to 12,000 characters; if they are longer, the groups that come last from A to Z are left out, and your Instructions say which, so you can ask an admin to shorten them. Instructions never widen what anyone may read or change: the assistant still works only as you, with your own access. Changes to the organization's and groups' instructions are in the audit trail, with who made them and how long they are, never their text.
Projects
In the SourceLace app, a project groups your chats about one piece of work. Its instructions (such as "Amounts in EUR, fiscal year starts in April") and its own skills apply to every chat in it, and its board keeps the charts, tables and files you pin from those chats. A project's page has a tab for each: Board, Chats, Skills, Sharing (for its owner) and About (its name and instructions). Long lists show 25 at a time, with Show more and a search box.
The Projects page
Projects in the quick links opens the list of every project you can open: your own, and those shared with you. Each row shows the project's name and instructions, how many of your chats are in it, when it was last active (its last change or your last chat in it), and whether you own it or it was shared with you (with what you may do). Find one with the search box (it looks in the name, the instructions and the owner), then narrow or reorder the list:
- Show: All, Mine, Shared with me or Pinned.
- Sort: Recently active (the default), Name A to Z, Newest first or Most chats.
The list loads a page at a time; More projects loads the next. New project opens a small form for the name of an empty project. Choosing Pinned from the sidebar's All pinned projects row opens this page already filtered.
Pinning a project
Pin the projects you come back to most. Press the star on a project's row, or Pin project on its page (once pinned, that button says Unpin project and shows a filled star); the project then appears under Pinned projects in the sidebar, just below Projects (up to five, the most recently pinned first; with more, All pinned projects opens the Projects page filtered to them). Press the star again, or Unpin project, to take it out. A pin is your own preference: nobody else sees which projects you pinned, and pinning changes nothing about the project itself, who can open it or its instructions. You can pin a project you own and one shared with you (up to 50 pins). If someone stops sharing a project with you, it leaves your pins; if its owner deletes it, it is gone from everyone's. A pin is not the same as pinning a chart or table to a project's board.
Starting a chat in a project
At the top of a project's page, Start a new chat in this project takes a question: press Enter (or Start chat) and a new chat opens in the project, already asking it, with the project's instructions and skills. Leave the box empty and press Start chat to open an empty chat in the project. Until your admin has set up the AI model the box is switched off. The other way still works: in any chat, Add to a project (or the folder beside it in the sidebar) moves it into one.
Starring chats, and chats in projects
The sidebar keeps your chats in order without any extra folders: projects are the folders. From the top it has New chat, the quick links (Search where your organization has Search everything, Getting started, Projects and Artifacts), then, if you pinned any, Pinned projects (up to five), then two panes: Navigation (a Recent projects section with three of your other projects, the most recently created or edited first, and, with more, an All N projects row; then the pages of the app, grouped into Work, Help and, for admins, a folded Administration group) and, below it, Chats (with Starred and Recent). Both panes show from the start. Each pane scrolls on its own, so loading older chats never pushes Navigation out of sight. On a short screen and on a phone the quick links sit at the top of Navigation and scroll with it, so Chats keeps room for your recent chats. On a phone the sidebar is a drawer behind the menu button; it closes when you choose something, including the page you are already on.
- Chats in a project are on that project's page: choose the project in Pinned projects, Recent projects (or Projects) to see them. They leave the Recent list, so it holds only chats that are in no project. If you would rather see them in both places, tick Show project chats in Recent in Your settings, under Chats; the setting is yours alone.
- Star a chat (the star beside it) to keep it in the Starred section at the top of Chats. Press the star again to unstar it. Stars are private to you.
- Move to project (the folder beside a chat) lists the projects you can open; type to find one. Remove from project puts the chat back in the main list. On a computer you can also drag a chat onto a project in Pinned projects or Recent projects (on a touch screen use Move to project, which reaches every project). You can move a chat into your own projects and into projects shared with you; chats stay private to you either way.
- Search chats looks through every chat, in projects or not, and shows which project each one is in. Where your organization does not have Search everything, Search chats in the quick links takes you to the box in Chats.
- Fold a section away by clicking its heading (Chats, Starred, Recent, Navigation, and the headings inside Navigation); click it again to show it. Fold Chats to give Navigation the whole sidebar, or fold Navigation to give your chats the whole sidebar. Admins open Administration to see its five groups: People and access, Connections, AI, usage and plan, Business setup, and Trust and health.
- Share the space between Navigation and Chats by dragging the bar between them up or down (or focus the bar and use the up and down arrow keys; double-click resets it). Navigation never takes more room than its links need.
- Show chats, navigation or both with the two buttons beside the SourceLace name (a speech bubble for Chats, three lines for Navigation). Both are on to begin with. Press one to hide that part, and the other takes the whole sidebar; press it again to bring it back. At least one part always stays on, so the button of the last one showing is greyed out.
- Hide the sidebar with the button beside the SourceLace name; a slim bar stays so you can show it again or start a new chat. Drag the sidebar's right edge to make it wider or narrower (or focus the edge and use the arrow keys; double-click resets it). SourceLace remembers the folded sections, the split and the width in your browser.
Sharing a project
Open the project and use Sharing to share it with a person (by email), a group, or everyone in your organization:
- Can use: they see the project under Shared with me, and its instructions and skills apply in their own chats in it.
- Can edit: they can also change the instructions and add, change or delete the project's skills. Only you can rename, share or delete the project.
Chats, pins and files stay private to each person, even in a shared project: nobody sees your chats or your board, and you do not see theirs. The assistant always works as the person asking, with their own logins and access. You can stop sharing at any time; people keep their chats, but the project's instructions and skills no longer apply to them. Projects can only be shared within your organization.
Kinds of chart
Ask for a chart in your own words and say the kind if you care. Bar charts compare categories; they can stand up (the default) or lie down: ask for a "horizontal bar chart" and the category names run down the left, flat and readable, with the first category at the top, which suits long names and rankings. Say "largest at the top" and the assistant orders them that way. Line charts show change over time, pie charts show parts of a whole (up to six slices), scatter charts show whether two numbers move together, and timeline charts show items from a start date to an end date. Click a bar, slice or dot to see the rows behind it, or press Show as table. SourceLace does not draw stacked bars, a second axis, maps or custom colours; the assistant says so rather than pretending. Example: "Show the ten accounts with the highest health score as a horizontal bar chart, largest at the top."
Refreshing a pinned chart
A pinned chart, table or file shows the numbers from when the assistant answered (the board says "As of" that time), with the check line of that answer under it ("Answer check: 1 thing to know", with Show); the line is not shown after a Refresh, because it describes the first read. Click Refresh to run the queries behind it again (every query, and the question across systems when it combined several) and redraw it with today's numbers. Refresh runs as you, with your own sign-ins and access, never as whoever first asked. The fresh rows are held briefly and expire like any result; the chat itself is not changed. Headline numbers the assistant worked out are left off a refreshed chart, since they came from the old rows.
If a refresh fails, for example because you are not connected to the source or no longer have access, the board says why and keeps showing the earlier picture with its date. Charts and tables pinned before Refresh existed show "Can't refresh: pinned before refresh existed"; ask the question again and pin the new one.
Artifacts: every file in one place
Artifacts (in the quick links, beside Projects) lists the files your chats and agents made in one place: the spreadsheets and CSV files you exported from a table, the Word, PowerPoint and PDF files the assistant made, the files an agent run made (kept in the agent's results chat), and the photos you attached to questions (shown as Uploaded in chat). Each row shows the file's name and type, where it came from (the chat or agent run, with a link, and the project if it is in one), who made it, when, its size, when it will be removed, and a Download button.
- Types. The buttons at the top show only spreadsheets, documents, presentations, PDFs, images or other, with how many of each. By default the list is grouped by type; untick Group by type for one plain list.
- Search finds files by name, by the title of their chat or by their project, among your newest 2,000 files. Source narrows to files made in chats or by agent runs. Sort by newest, oldest, removed soonest, name, type or size. The list shows 50 files at a time.
- Whose. Mine is every file of your own chats. Shared in my projects is only the files of your chats that are in a project someone shared with you. Sharing a project shares its instructions and skills, never anyone's chats or files, so you never see other people's files here, and nobody at SourceLace sees them either.
- When files are removed. A file stays with its chat. It is removed when the chat is: your organization's retention period after the last activity in that chat (30 days unless your admin chose otherwise). Ask a new question in the chat, or download the file, before the day shown. Files in a chat you delete go at once. Starting a conversation in the chat again moves the date.
Nothing new is kept for this page: it lists the files the chats already keep, encrypted with your organization's key.
Spreadsheets of your results
Ask for the data itself, such as "Give me every renewal at risk as a spreadsheet", and the assistant makes an Excel (xlsx) or CSV file of the results it found, with a download card under its reply.
- Every row, not a sample: up to 100,000 rows on each sheet, and all of the result's columns unless you ask for some. One Excel file can hold several sheets, one for each result; a CSV file holds one table.
- An About sheet comes first in Excel files: what the file is, when and from which systems it was made, how many rows each sheet holds, and whether any rows were left out because a result was larger than a sheet allows.
- The About sheet also says what SourceLace checked. At its end, under About this answer, it repeats what the line under the answer said: the headline (for example "Answer check: 1 thing to know"), each error and thing to know with what you can do about it, and the lists behind Show. A CSV file has no such note, because an extra row would break the file for the programs that read it; the line under the answer is where its checks are.
- Data protection still applies: values your organization masks are masked in the file too, and hidden fields are not in it at all.
- Nothing is read again: the file is made from the results the assistant already has, so it shows the same data as the answer. Ask the question again for newer numbers.
- Safe to open: text from your systems is always written as text, so a value that looks like a spreadsheet formula never runs as one.
The file is kept with the chat and deleted with it. Agents with a Make a document action can make spreadsheets too; they go to the agent's results project.
Tables in answers
To answer one question, the assistant often reads several tables first, for example accounts from your CRM and invoices from your ERP, and then combines them. A chat answer shows you its own table, the one that answers your question. The tables it was read from are folded under it, one click away.
Answer only (the usual). You ask: "Which accounts with open renewals have overdue invoices?" You see one table, the answer, with a caption such as "14 rows from Salesforce and NetSuite". Under it is a closed fold, Built from 2 tables. Nothing is lost: it is only folded.
More than one table. When you ask for two tables ("the accounts, and then their contacts"), you normally get both, one after the other in the reply, in the order you asked. The assistant marks each table it makes as one you asked for or as a helper (a count, a lookup); the helpers fold under Built from, so an extra column or a count it needed on the way never takes the place of your table. A table the assistant only read on the way to another one, such as raw rows it then reshaped into your table, also folds under Built from, even if it was first marked as one you asked for; a reply with just one table always shows it. If you asked for certain columns in a certain order, the assistant aims for exactly those; check the table. If a table does not look like what you asked for, say so in the chat and ask again with the columns you want.
A table built the way you asked
When you ask for particular columns, names or an order ("the contacts with Full name and Job title, sorted by name, the top 8"), the assistant tells SourceLace what the table should look like, and SourceLace builds the table itself from the rows it read, the same way for every system. Under the table, one line says it was built by SourceLace from a plan and puts the plan in words: each column next to the heading it was given, how it was sorted, how many rows it kept, and, for counts and totals, how they were grouped. A heading the assistant chose is shown next to the system's own name for the column ("Discount shown as Revenue"), so a renamed column is never hidden. If SourceLace could not build the table the assistant asked for, for example because a column it named is not in the rows, you see the rows as they were read and the line under the answer says that the plan was refused and why. The plan is the assistant's reading of your question: SourceLace builds the table that the plan describes, and does not check that the plan is what you meant, so compare the headings and the order with your question. Totals over a result that was cut short are only built when the assistant says so, and the line under the answer then says how many rows they cover.
When an earlier table kept only its top rows. If a table was cut to its first rows (for example "the top 8", or a question that carried its own limit) and you then build another table from it, SourceLace tells you in the line under that table, as a thing to know (never an error). A total says: "This total is built only from the first 8 rows of an earlier table, not all of your data." A table in another order, with a new limit, after a filter or joined says: "This table is built only from the first 8 rows of an earlier table, not all of your data." This holds even when nothing is wrong with the numbers: the top 2 of the top 8 is still the top 2, and the line still says where it came from. A table that only shows some of the same columns of those rows says so calmly under Checked by SourceLace: "Shows rows of an earlier table that kept only its first 8 rows." Next to either of the first two lines, choose Redo on all rows: it adds this message to the same chat as your own words, "Do that again on all rows, read fresh from the source.", and the assistant answers again below, reading the rows again from the system, within your limits and read rules, instead of working from the cut table. It is shown only under the latest answer, so it always redoes that one. It reads the system again, so it takes as long as a new question and counts as one; it uses your own access and every limit, cap and rule that applies to your questions, and nothing in your chat is deleted. You cannot press it again while an answer is being written. It is for chats only, not for boards. It is not Ask live, which reads a saved answer again from the system.
Opening what it was built from.
- Click Built from 2 tables. It opens a short list: "Table 1 of 2: 212 rows from Salesforce" and "Table 2 of 2: 1,940 rows from NetSuite".
- Click one to read its rows, or click Expand all, next to the line (it shows when there is more than one), to open both. The button then says Collapse all.
- Every table you open has the same controls as the answer's table, and its own Where this came from buttons. You can still pin it to a project's board and see How it was read. The label above an open table says which one it is ("Table 1 of 2: 212 rows from Salesforce").
What you open lasts while you stay on the page. SourceLace does not keep it, because the column names in an answer come from your systems; after a reload the folds are closed again.
The defaults. Under Your settings → Chats, Tables in answers chooses what every answer starts with, and it is only a default. Besides Answer only, Folded folds just the tables that were combined into a later table, and All tables shows every table the assistant read, in full. The most specific choice wins:
- Yours. Choose under Your settings → Chats. It applies to all your chats. Pick the first option, which names the project's or organization's choice, to follow it again.
- The project's. A project's owner, and people it is shared with to edit, choose on the project's About tab. It applies in the project's chats to everyone who has not chosen for themselves.
- The organization's. Admins choose under Chat defaults → Chats → Tables in answers. It applies everywhere else. Answer only is the default for anyone who has never saved this setting; an organization or project that already saved a choice keeps it until someone changes it.
Folding only changes what the chat shows, not what the assistant reads or what is recorded in the audit trail.
The line above a table. Each table has one line of controls above it. The search box is at the left (it shows when the table has more than 5 rows). At the right are Show where columns came from, Ask again with record ids or Show record ids, Columns, Fit columns, Copy table, Export (a small menu: CSV file or Excel file; for a table that came from your systems, not one the assistant typed) and, last, Expand; the ones that need explaining have a small ⓘ. On a phone the line would wrap, so everything but Copy table is behind one More button, Export with them.
Working with a table. Every table in a chat answer can be set up the way you like: drag the edge of a column heading (or select it and press the left and right arrow keys) to make the column wider or narrower, and use Columns above the table to move a column left or right. Long text is cut to two lines; Show full text opens it and Show less closes it. The header row stays in view while you scroll, and a table wider than the page scrolls sideways inside its own box. Your changes last while you stay on the page and are not remembered after a reload. Columns can be moved but not hidden, so an answer never looks as if a column is missing. Wider tables and fitted columns. An answer's table is as wide as its columns need, never narrower than the text, and wider than the text where the page has room, up to the width of the chat less a margin or about 1,600 pixels, and it is centred under the text, so it grows equally to both sides (on a phone, rows stay cards). Hiding the sidebar gives the chat, and so the text and the tables, more room. Each column starts as wide as its heading and its values need, so a number does not wrap in the middle. Fit columns above the table sets every column back to that width after you have dragged some; your own widths win until you press it. A table the assistant typed in its reply works the same way in the chat (small ones show only Copy table).
Numbers sit at the right. A column of numbers is lined up at the right, its heading too, so digits of the same place sit under each other. A column whose heading ends with the word id, code, number, no, zip, postal, phone, version or ref ("Customer number", "Postal code"), and any value written with leading zeros (00017), stays at the left, because those are names and not quantities. The heading is only a hint: a column of ids with no such word in its heading is read from its values alone, and "Number of orders" is a quantity. Only the position changes; the value you copy or download is exactly what your system sent.
Numbers. When every value in a column is a long decimal, such as 591630.0000000000 and 1500000.5000000000, the column is shown tidier: the zeros all its values end with are dropped and thousands are separated (591,630.0 and 1,500,000.5). The separators are always a comma for thousands and a dot for decimals, whatever your browser's language. Anything else is shown exactly as your system sent it: whole numbers, customer numbers, ids, postcodes, dates, versions, codes, money with two or four decimals, any column that mixes kinds of value, a text column whose heading ends with the word id, code, number, no, zip, postal, phone, version or ref, and all text. Copy table, spreadsheets and the text that appears when you hover a number keep the exact value; a screen reader or a touch screen (no hover) gets the tidied value, so use Copy table where the exact digits matter. A value your organization hides is never shown here.
The steps the assistant took. While an answer is being written, each step shows as it happens ("Searching bigquery...", "Querying salesforce:dev"). Once the answer is finished, the steps fold under one line, Worked in 12 steps. Click it to see them all; each step still has its own Details. If you opened a step's Details, or had scrolled up, while the answer was written, the list stays open when it ends. A step that failed and that the answer did not get past stays in view under the line, in red, and the line says so: Worked in 12 steps (1 had a problem, shown below). Which steps you opened is not remembered; after a reload the line is closed again.
When a step did not work, and when it was corrected. The assistant works in steps, and a step can fail, for example when a query names a column the system does not have. Often the assistant simply tries again with a better query, and the answer is right. The line above the steps then says so: Worked in 25 steps (1 step corrected along the way). Open it and the step that failed is in its place in the list, as a calm grey note (no red, no alert) that says in plain words why it failed ("The query asked for a column or table that does not exist.") and The assistant corrected this and carried on. The system's own message sits behind Details. Corrected is said only for a mistake the assistant can fix by asking again (a query that named a column that does not exist, a query that was not valid, a timeout, or the system's own trouble) when a later query of the same kind on the same system worked in the same answer; it does not mean the answer is wrong, and the page cannot tell a rewritten query from a different query on the same system, so check what the answer rests on. A failure that nothing fixed stays red, even if the assistant then wrote an answer, and so does any failure that is not clearly one of those mistakes, and so do a refused sign-in or permission, a limit, a stopped answer, a setting your admin must change, a refusal by data protection and any failed change or file, even when something else worked afterwards. Chats from before this was added show every failed step in red: the line says (1 had a problem, shown below), and the red step shows under the line with what it means and what to do, with Details behind it. Details shows the system's message with anything that looks like a sign-in removed, as far as that can be done; it is shown only to you.
Your own AI app (MCP) shows tables its own way; these choices are for the SourceLace app.
Making a table or chart wider
Press Expand (last in the line above a table, or in the corner of a chart beside Show as table) to make that one table or chart as wide as the chat, so you can see more columns or a long chart without scrolling sideways. The rest of the page stays where it is. Press Collapse, or the Escape key while you are in that table or chart, to put it back; the keyboard stays on the button. Escape pressed anywhere else, such as in the box you type your question in, does not touch it. Expand only shows where there is room to make the block wider, so a narrow window may not show it. Expand is not remembered: when you reload the page or come back later, every table and chart starts at its usual width. On a phone there is no Expand, because tables and charts already use the full width. Example: a table with 12 columns that does not fit beside the text: press Expand to see all of them. Charts are drawn at a comfortable width, up to about 1,000 pixels, whatever the width of a table.
When an answer fails
If SourceLace cannot finish an answer, a red message says what happened and what to do next. Press Try again to ask the same question again. Press Copy details for your admin to copy a short note (when it happened, which chat, and the message you were shown, with no question or answer text) that you can paste into a message to your admin; nothing is sent until you paste it. The button is not shown if your organization turned the Copy buttons off. To change the last question and ask it again, press Edit and resend under it: its words go back in the box, and the earlier question and answer stay in the chat. On a phone, a row's card shows its first few fields and Show all fields, at the end of the card, opens the rest; in the menu, a chat's ⋯ button holds Move to a project, Rename and Delete.
Sorting and searching a table
Every table in a chat answer can be sorted and searched. Both work on the rows you can see on the screen, so they are instant and send nothing anywhere. A worked example: you ask for your biggest customers and the answer shows 50 rows, with the line "2,000 rows from Salesforce (more rows not shown)" under it.
- Sort. Click the name of the Revenue column. The 50 rows are put in order from smallest to largest, an arrow shows the direction, and the line under the table now also says "Sorted these 50 of 2,000 rows". Click again for largest first, and a third time to go back to the order the assistant gave. One column is sorted at a time. On a phone, where rows are cards, use Sort by above the cards and choose the column and the direction.
- Search. Above a table of more than 5 rows, type in Search these rows, for example
acme. Only the rows with that word in any column stay, the matches are highlighted, and the line under the table says "3 of 50 rows match". Capital letters and accents do not matter, soeclairfindsÉclair. Several words must all appear:acme berlinkeeps rows that have both. A grouped number is found by its digits too, so1234567finds1,234,567. Clear (or Escape) shows every row again. If nothing matches, the table says "No rows match". - Both together. A filter you set from a chart comes first, then the search, then the sort.
What "Sorted these 50 of 2,000 rows" means. Sorting puts the 50 rows on your screen in order. It does not look at the other 1,950, so the first row is the biggest of those 50, not necessarily the biggest of all 2,000. For the real top 10, ask the assistant ("show the 10 customers with the highest revenue"): it sorts the whole result before it answers. When the assistant shows the next rows of a result ("Show next rows"), each set of rows is its own table with its own sort and search, and the line says which rows it applies to, for example "Sorted rows 51 to 100 of 2,000".
How values are compared. Numbers are compared as numbers (9 comes before 10), dates by time, and text in natural order that ignores capitals and accents ("Item 2" before "Item 10"). A column of ids with leading zeros, such as 00017, is sorted as text, so the zeros are kept. Empty values always come last, whichever way you sort. Rows that are equal keep the order they had. What decides is the value as your system sent it, not the tidied number on the screen. Only plain numbers (1234.5, -7) sort as numbers; 5., 1e3, 1,234.5, $5 and 50% are text. Only ISO dates and times (2026-11-02, 2026-11-02T09:30:00Z) sort by time; a date written 03/02/2026 is text. A date with no time and a time with no zone are both read as UTC, and an impossible date such as 2026-02-30 is text. In a column of numbers, a value such as N/A or - goes after the numbers. Whether a column counts as numbers, dates or text is decided from the rows of that table, so two sets of rows of one result can sort differently, for example when one set has a value such as n/a in a column of numbers.
Copy table copies what you see. After a search and a sort, Copy table copies only the rows that are shown, in the order shown, with each value exactly as your system sent it. Your choices last while you stay on the page and are not remembered after a reload. Tables the assistant typed in its reply can be sorted and searched the same way.
What a file says about the answer
A Word, PowerPoint or PDF file the assistant makes ends with About this answer (in PowerPoint, on as many closing slides as it needs), and an Excel file carries the same on its About sheet. It says what the line under the answer said when the file was made: the headline, each error and thing to know with its tip, and the lists behind Show, so a file that went to someone else does not look more complete than the answer was. A chart drawn from a table that was cut carries one short note saying so. If the file is made from tables of more than one answer, it has one note for each answer ("Answer 1 of 2"). A file made when SourceLace holds no check for the answer (an agent's file, or an answer saved before the line existed) says "SourceLace's checks are not available for this answer." instead; it never says "all clear" without one. The note holds no values from your tables; the columns it names are only the ones the line under the answer names too. A CSV file carries none.
Exporting a table
A table in an answer shows the first rows of a result that can hold many more. Export makes a file of all the rows of that result, up to a limit, and keeps it with the chat, so you can download it again. You choose the kind: a CSV file (plain text that opens in Excel, Google Sheets and many other programs) or an Excel file (an .xlsx workbook with a first sheet that says what is in it and a sheet called Results with a bold heading row). It is made the way the files the assistant makes for you are made. A worked example: you ask for open invoices, the answer shows 50 rows and says "2,000 rows".
- Open Export above the table (on a phone, press More first) and choose CSV file or Excel file. The button and a line above the table say "Preparing…" while SourceLace makes the file, and the button cannot be pressed twice. Hide in that line only hides the line: SourceLace still makes the file and keeps it with the chat, so its card can still appear in the chat afterwards, and the button keeps saying "Preparing…" until it is done.
- The file appears as a card above the table, and also as a new message at the end of the chat, like the files the assistant makes: its name (for example
sourcelace-results-20261113-0930.csv; the date and time, in UTC, are in the name and nothing from your question or your systems), its size, and a line saying how many rows it holds, such as "Your CSV file is ready: 2,000 rows." - Press Download on a card. Your browser saves the file. The card in the chat stays with the chat: leave and come back, and it is still there.
- Open it in Excel. All 2,000 rows are there, in the order the assistant found them, whatever you sorted or searched on screen.
A very large result, or one with very many columns, is limited to the first rows that fit (at most 100,000 rows and 2,000,000 cells, and 15 MB for the file), and the line says so ("the first 100,000 of 340,000 rows"); an Excel file's first sheet says it too. An Excel file's About sheet also ends with About this answer: what the line under the answer said (what a file says about the answer); a CSV file carries none. A file holds at most 200 columns: for a table with more, the line says "Only the first 200 of its 250 columns are in the file". If the answer was already marked Partial data, the file has the same gap.
What is in the file. The rows the assistant sees: masked values stay masked in the file and columns your organization hides are not in it. Text is saved as text, so a value that starts with =, +, - or @ cannot run as a formula (a CSV file puts an apostrophe in front of it). An Excel file keeps numbers as numbers; a very long text value (over 32,767 characters, Excel's limit) is cut there and the first sheet says how many were cut; a CSV file keeps it whole. The file is kept with the chat like every other file SourceLace makes: encrypted with your organization's key, and deleted with the chat. Your administrator can see in the audit log that you exported a result, and how many rows, but not what was in it.
Sorting and searching do not change it. The file is the whole result, not the rows on your screen. To get only some rows, ask the assistant to run the question again with a filter, such as "only the invoices from Berlin", and export that table. Results are held for a limited time, so export soon after asking. Each time you choose CSV file or Excel file a new file is made, and your administrator may limit how many exports one person can make in an hour; an export that fails after SourceLace started making the file still counts toward that limit. If you leave the chat while a file is being made, the file may still appear with the chat.
When something goes wrong
| What you see | What it means | What to do |
|---|---|---|
| "Exports are busy right now. Try again in about 30 seconds." | Your organization already has the most exports running at once. | Wait the time said and choose again. |
| "Exports are busy right now. Try again in about 5 seconds." | SourceLace is making other Excel files at this moment. | Wait a few seconds and choose again; a CSV file may work sooner. |
| "You have made the most exports allowed for now." | You reached the number of exports allowed in an hour. | Try again after the time said. |
| "This result is no longer held, or is not in this chat." | Results are held for a limited time, and this one has gone, or the table is from another chat. | Ask the question again and export the new table. |
| "This result is too large to export in one file." | The file would be over 15 MB. | Ask the question again with fewer rows or columns. |
| "A reply is still being written in this chat." | The assistant is still answering. | Wait for the answer to finish, then export. |
| "You can't export this result right now." | Your access to one of its systems has changed. | Ask your admin, or ask the question again. |
| "SourceLace is busy or restarting." | The server is changing. | Try again in a moment. |
| "You're signed out." | Your sign-in ended. | Sign in again, then export. |
| "Can't reach SourceLace." | Your connection broke before SourceLace answered. | Check your connection and choose again. |
| "The export could not be completed." | Something unexpected went wrong. | Choose again; if it keeps happening, tell your admin. |
| "This chat could not be found." | The chat was deleted. | Reload the page. |
| The browser blocks the download | Your browser's download settings are strict. | Press Download on the card again, or allow downloads from SourceLace in your browser. |
| One column in Excel, all text in one cell (CSV) | Excel read the file with the wrong list separator. | Use Data, From Text/CSV and choose the comma, or export an Excel file. |
| A value starts with an apostrophe (CSV) | The value began with =, +, - or @; the apostrophe keeps it from running as a formula. |
Delete the apostrophe if you need the plain text. |
Where this came from
Every column of an answer that SourceLace can trace has a small button next to its name. It says which field in which system the column's values came from.
One column.
- Find the column's heading. Next to its name there is a small button showing the logo of the system the column was read from ("fx" for a computed column, the system's letters when SourceLace has no logo for it); its name for a screen reader is "Where Name came from". On a phone, where each row is a card, the column names are listed in a line above the cards, each with its button.
- Click it. A small panel titled Where this came from opens: "From Salesforce: Account.Name".
- For a total such as
revenue, the button looks different (it shows "fx") and the panel says "Computed from NetSuite: Transaction.amount": the value is made from that field, not copied from it. - Press Escape, click elsewhere or click the button again to close it. On a phone the panel opens across the bottom of the screen.
A column joined from another system. When an answer joins systems, a column read from the other system can be empty for an account for two reasons: nothing in that system matched the account, or the matched record has nothing in that field. In that column an empty cell shows a dash (a short horizontal line, read out as "no match, or empty in the system" by a screen reader) and a line under the table says what it means; a real empty value elsewhere stays blank, and a count of zero shows 0. The column's Where this came from panel then also says how the rows were matched, for example "Matched by comparing Snowflake ACCOUNTS.NAME and Salesforce Account.Name", and warns when that match depends on another one. A dash means the question did not find a matching row, not that the system holds no data for the account: it can simply mean the names are spelled differently. Copy table, search, sort and exports treat the cell as empty.
Every column at once. In the line above the table (on a phone, behind More), press Show where columns came from. Under every heading, the system and field now show ("Salesforce Account.Name", "NetSuite Transaction.amount (computed)"). Press it again to hide them. It only changes this table.
Record ids. A record id is the id a row has in its system: "Account 001A0" tells you which record in Salesforce a row came from. A table only has ids if the assistant fetched them.
- If the table has them, the toolbar shows Show record ids. Press it and a Record ids column appears at the end.
- If it does not, the toolbar shows Ask again with record ids. Press it and SourceLace sends your last question again as your next message, asking for each record's id in the table that answers it. The assistant is asked to include the ids; when it does, the new table's toolbar offers Show record ids. The button belongs to the latest answer's own table and waits while an answer is being written.
- A total has no ids to show, because it comes from many records. Ask for the records themselves, such as "list the accounts with their ids".
What a table starts with. Under Your settings → Chats, Where this came from chooses what each table starts with; a project's About tab and an admin's Chat defaults → Chats choose it too, the most specific winning, as for tables. On request (the usual) starts with the buttons only. Open by default starts every table with the system and field under each column's name. Where your organization uses matching rules, Open by default, and ask for record ids in future answers also asks the assistant to fetch each record's id in answers you ask from now on, and starts with the record ids shown; it does not change answers already on the page.
When it does not show. If SourceLace cannot tell where a column came from, the table has no buttons (for an answer across systems it also says so under the table: "Where these columns came from can't be shown for this answer."). It never guesses. Some answers cannot say where a column came from, for example when the question reads fields through a related record (such as Account.Name in a Salesforce question), joins several tables in one statement, reads files or mail, or when the assistant computed a column in a way SourceLace cannot read back to a field. Ask the assistant for the fields directly to get the buttons.
A line instead of a copied table. Sometimes the assistant types the rows of an answer again as its own table. That copy has no copy button, fold or where-this-came-from buttons, so when it is an exact copy of the answer's table SourceLace replaces it with the line "(These rows are in this answer's result table.)". The real table is part of the same answer (under Built from when it is folded). A table the assistant made itself, such as totals, a ranking or a comparison, stays.
The line under an answer
Under each answer in the app, SourceLace adds one short line that it writes itself, not the assistant. It starts with Answer check: and says what SourceLace found, with an icon of its own for each case (a check, a triangle or a cross): Answer check: all clear, Answer check: 2 things to know, Answer check: 1 error, or Answer check: 1 error, 2 things to know.
- An error means something the answer relies on failed and nothing put it right, for example a step that failed and no later step read the same thing. An answer with an error starts with one short sentence saying so.
- A thing to know is not a failure. Your question ran fine, and the answer is limited in a way you should know: rows were cut short, rows came from saved answers, a system you chose was not read, a column is blank because its rows were not found in another system, some rows have no match in a system, a count of 0 may be wrong because a system was read only in part, or a column held values that are not numbers (such as NaN) and its figures were worked out without them; the note gives how many were left out, and the assistant's own figures say so too. A blank column is explained right under its table, for example "Your question ran fine. These columns are blank because the rows were not found in Snowflake.", and when several systems are named each is shown with its own columns. When some rows, not all, have no match it says how many, for example "Your question ran fine. 7 of these 10 rows have no match in Snowflake." Each counts once in the line. A count of rows that found nothing shows 0; SourceLace only says something about it when the system it counts was read in part ("A count of 0 here may be wrong: SourceLace read only part of Snowflake.").
- All clear means every check SourceLace makes passed and there is nothing to know. A step that failed and was then redone is listed under Checked by SourceLace, so you can see it.
If the assistant tried again and the second table fills in the blank columns of the first, the first stays in the answer under Built from, labelled Earlier attempt, replaced by Table 2 (or ...by the answer table). Its blank columns are then listed under Checked by SourceLace, not counted in the line, so the line is clear when the final table is right. SourceLace decides this itself, never from what the assistant says, and it never changes or removes the first table.
When an answer shows more than one table, each is labelled Table 1, Table 2 and so on, and the lines behind Show say which table they are about. The tables folded under Built from are named Input table 1 of 3 and so on; a line that is about the answer's own table and one of those says "this table and Input table 1 of 2". In the All tables view, the answer's own table is headed Answer table, and the lines say "Answer table" instead of "this table".
Click Show to open the line. You see the errors, then Good to know, most of them with a tip for what to try (a few, such as a refused or stopped answer or a generic failure, have none). Then comes Checked by SourceLace: how the answer ended, whether the steps worked, whether rows were cut, when each system was read, and which columns your organization masks ("Masked by your organization's rules:" and the columns; press the ⓘ beside it to read what that means). A masked column is a fact, so it is listed there and never counted in the line. Last is Chosen by the assistant, not checked by SourceLace: which systems, objects and fields the assistant chose to read, the filters it used and what it took words like "open" or "this quarter" to mean, how records were matched across systems, and the wording of its text. SourceLace reports what it can show about how the answer was made: steps that failed and whether a later step redid them, results that were cut and why, joins that matched nothing, values that are not numbers and were left out of a figure, data from saved answers and how old it is, and fields your organization protects. It does not check that the assistant chose the right data, or that the numbers in its text are right, so read that list too.
Times are shown in your time zone. Copy answer and Read aloud include the line, the sentence at the top of an answer with an error, and each error and thing to know with its tip, where it has one; for a blank column they also say which column (and, with more than one table, which table), for example "...not found in Snowflake: "Tier" (Table 1)."; with several systems each is said with its own columns; Read aloud says "the column Tier". If you stop an answer, its line appears when you reopen the chat. Answers made before the line existed do not have one, and an answer saved by an older version of SourceLace says "This answer was saved by an older version of SourceLace" in its place. The Answered from line is not part of Copy answer or Read aloud.
A note about an earlier table that kept only its first rows has a Redo on all rows button beside it (see A table built the way you asked).
Only SourceLace can write this line. If the assistant's text imitates it, that is itself reported as a thing to know. The line names the systems, objects and masked columns of the answer, and columns with values left out of their figures, so it is kept only inside the chat, encrypted like the chat and deleted with it after your organization's chat retention period; it is never kept in your browser.
Where your next answer starts
When your last answer has a table that SourceLace is still holding (SourceLace keeps results for a short time), a small switch appears beside the send box with two choices. Use the data already read is the usual one: your next question works from the table already shown, which is quick, and your systems are not asked again. Read the systems again asks them again for that one question, so the answer shows their current data. It applies to the next question only and goes back to the usual choice after you send it; SourceLace does not remember it. For example, after "open invoices by region", the question "just EMEA" can use the table already shown, or you can choose Read the systems again to see the current figures. The switch is not shown for your first question, after an answer with no table, or when the held table is gone. Redo on all rows stays on the answer, as before.
Where an answer came from
Press Show under the answer's check line: the first line under Checked by SourceLace starts with Answered from and says what the answer used: the systems it read just now ("Answered from: Snowflake and Salesforce, read now."), the table held from your earlier answer ("Answered from: the table held from your earlier answer (read from Snowflake and Salesforce at 14:05)."; when the table was read on another day the date is shown too), or saved answers with their age ("Answered from: saved answers, read from Snowflake 2 days ago."). An answer that used more than one joins them with a semicolon. Answers saved before this line existed do not have it.
Searching and sorting lists
Every long list in the app works the same way: Agents, Skills, People, Groups, Shared agents, Audit log, Invoices and the rest. Above the list there is a search box; type a few words and only the rows that contain all of them stay (capital letters and accents do not matter). Click a column's name to sort by it: once for A to Z (or smallest to largest), again for the other way, and, on lists that hold everything on the page, a third time to go back to the original order. Numbers, money, percentages and dates sort as what they are (in whatever language and number format your browser uses), a status sorts by the word shown, and empty cells always go last. On a phone, use Sort by above the list. Copy table copies the rows you see, in the order you see them.
On a list that loads a page at a time (Agents, Skills, People, Groups, Shared agents, Runs, Change rules, Matching rules, Read rules, Message queues, Audit log, Usage and the others), the search and the sort cover every row of the list, not only the page on screen. A few columns do not sort (for example the Result of a run, which the database cannot order quickly); their headings are plain text. The Audit log sorts by When: click it for oldest first, again for newest first. SourceLace does not save what you type or the order you choose. One exception: when you export the Audit log, the filters you used, including a search word, are written to the audit trail like any other entry (it is encrypted and kept for your plan's audit history). What happens to the words you type depends on the list. On People, an agent's Runs list, a shared agent's Subscribers and Runs, the Audit log (its search and its Person, Action, Source and Object choices, and what you type to find them) and the chat and project searches in the sidebar, which can hold people's names and email addresses, the words and choices stay in the page and are sent to SourceLace to do the search without being put in the page's address, so they do not appear in a link you copy or in your browser's history, and a refresh or the Back button starts with an empty search box and no person chosen (the period, the order you chose and the page may stay in the address). On the other lists that have filters above them (Agents, Skills, Groups and similar), the words are kept in the page's address so that Back and a copied link return to the same view: they are sent to SourceLace as part of the address, and they can show in your browser history, in a link you share and in the records of SourceLace's servers, which keep addresses for 365 days; do not type anything private into those searches. On every other list the words reset when you leave the page.
Choosing the AI model for your chats
If your organization offers more than one AI model, Model, under Options below the message box in any chat, new or old, lets you choose the one that answers your questions. When you have chosen your own, its name shows beside Options. It applies to all your chats, and only to you. Pick the first option, which names your organization's model, to follow it again.
| Model | Best for |
|---|---|
| Sonnet (the usual) | Most questions: a good balance of answer quality and cost. |
| Opus | Hard questions that need the most thorough answer. It is larger, so each question uses more of your organization's credits. |
| Haiku | Quick, simple questions. It is smaller and faster and uses fewer credits. |
Your admin decides which models your organization offers, and on the included AI your plan may offer only some of them. If a model you chose is no longer offered, your chats follow your organization's model. Search everything and AI agents always use your organization's model.
How full is this chat? (the context meter)
Every question you ask sends the AI model the whole chat so far: your questions, its answers and the results it read from your systems. A model can read only so much at once (its window), so a long chat slowly fills it. The gauge at the top of the chat, beside Search everything and New chat, shows how full this chat is. Select it to open the Context panel; on a phone it opens from the bottom. It never moves the chat.
What the panel shows.
- How full: for example "26% full, 52k of 200k tokens". A token is about three-quarters of a word. The bar is split by what fills the chat. The gauge turns orange at 80%. If your AI model's window is not known (a model your admin typed in under their own provider account), the panel shows the tokens used and no percent.
- What fills it: SourceLace's own instructions and tools (always sent, about 9,000 tokens); what your organization set (standing instructions, the project's, the skills list and your role guide; ask your admin to change them); updates SourceLace was told during the chat; your questions; photos you attached; SourceLace's answers; the AI's earlier reasoning (Claude models only); the results read from your systems (up to 50 rows each, usually the biggest part); and skills opened in this chat. Large results that SourceLace holds and analyzes itself are not counted: the panel says Held outside the chat.
- The sizes are estimates worked out from the amount of text and scaled to the exact total the AI provider reports for your last question. After you change something, the total is an estimate until your next question. The panel never shows names from your systems: a result is listed as, for example, "Result of question 2, 50 rows, about 14k".
- The cache note: the AI provider keeps a chat in its memory for a short time so the next question is faster and cheaper. The panel says when that runs out and about what your next question costs in credits while it is kept and after. Anything that changes what is sent makes the provider read the chat again once.
What you can do about a full chat. Each button shows a preview first (what it will do, about how much it frees, and whether it uses an AI call), and nothing changes until you press Apply. Afterwards Undo brings back what was sent before. Your saved chat always keeps every message: only what is sent to the AI changes.
| Button | What it does | Uses an AI call? |
|---|---|---|
| Compact all | Summarizes everything older than your last questions (you choose how many to keep word for word, 1 to 4, usually 2), and drops the AI's earlier reasoning and the text of skills opened earlier. Your latest results and the questions you keep are never touched. | Yes, one small one, counted in your organization's credits like a question |
| Forget big results | Stops sending old results from your systems to the AI, all but the latest. They stay in the chat; ask again if the AI needs the rows. | No |
| New chat from summary | Opens a new chat that begins with a short summary of this one. This chat stays as it is. | Yes, one small one |
You can also open a line, such as Results read from your systems, and forget or drop one item at a time. A summary covers questions and answers together, so they are compacted by position, not one by one. When a summary is in use the chat shows Earlier conversation was summarized, with Undo.
Compacting automatically
When a chat is nearly full it would otherwise stop with a message to start a new chat. So, unless your admin or you turned it off, SourceLace compacts it before your next question is sent, only when it is nearly full:
- At 85% of the model's window it first forgets old big results (no AI call; the results of your last 3 questions stay).
- If it is still above 70%, it summarizes everything older than your last 3 questions.
It never touches the results, the AI's reasoning or the opened-skill text of your last 3 answered questions, only older ones (a question the AI refused or did not answer does not count among them). Between 85% and 95% it waits for the provider's memory of the chat to run out (a rewrite then costs nothing extra), and above 95% it does not wait. It does this at most once an hour for a chat (the one exception: when the AI has just refused a chat as too long it tries once at once, then goes back to the hour), and shows each change in the chat (Earlier conversation was summarized automatically, Older results were left out...) with Undo; the Context panel lists every change in force under Changes in force, each with its own Undo. If the summary cannot be written, your question is still answered. For a model whose window is not known, it measures against 100,000 tokens.
The switch is in the Context panel, Compact automatically, and is yours. Your admin can switch it off for the whole organization (see Full chats under Chat defaults in the admin guide); then your switch is off too and the panel says "turned off by your admin".
When a message was sent
Hover over a message (or focus it with the keyboard, or tap it on a phone) to see how long ago it was sent, such as "3 minutes ago", and its full date and time. When you come back to a chat after a long pause, a line such as "Resumed after 2 hours 40 minutes" appears before your next question. When the pause was longer than the AI provider keeps the chat in its memory, the line says the next answer read the chat again at the full price.
Asking about SourceLace itself
This manual is built into SourceLace. Ask the SourceLace assistant (or your own AI app, if you connected one) questions such as "How do I connect Snowflake?" or "How long does SourceLace keep my data?": it searches this manual with the search_docs tool and answers with a link to the page. You can also download any page, or the whole manual, as a PDF from the docs site. For short answers by topic, see the frequently asked questions.
Getting help to use SourceLace
You can also ask the assistant to help you do something in SourceLace, such as "Help me set up an agent and test it" or "What should I do next to finish setup?". It asks what you want to achieve, gives you a short plan, and takes you through it one step at a time: what each step does, what to do, and how you will know it worked. For each step that happens on a screen it shows an Open button that takes you there. It reads your own Getting started progress, so it can tell you what is already done. It only suggests screens you may open: if a step needs an admin, it says so. The assistant never changes settings, adds people or asks for passwords or keys in the chat: you make each change yourself on its screen.
Which systems SourceLace connects to
See Connecting your systems for the full list, who sets each one up, and what people can read and change in each.