Your workspace's sources

Your workspace's sources

Every workspace keeps a short list of where its data lives: the spreadsheets and documents you uploaded, Google Sheets and other links you pasted, the reports the assistant built for you, the websites it logs into, and the integrations connected to it. The assistant reads that list at the start of every chat, so when you ask "how are sales this month?" it opens the right table — and doesn't answer "there's no data" while the report it built last week sits in the folder.

What gets added, and when

You don't have to do anything. A source is added when:

• you upload a file — a spreadsheet, CSV, PDF, document;
• you paste a link in a message — a Google Sheet or Doc, or any other web address;
• the assistant builds a report for you — it remembers which files it was built from;
• the assistant logs into a website in its browser — only the site's address is kept;
• an integration is connected to the workspace;
• a data table is created in the workspace's app data — by the assistant, by you through the public API, or by a template. The workspace's own service tables (the support inbox's drafts and statistics, the inbox, briefs and tasks of the «Personal assistant» template) are not listed.

The assistant can also add one itself when you tell it about a source ("our price list is on the site, under /prices").

What is remembered — and what never is

For each source the list keeps what it is, whom or what it is about ("по: Арман"), when to use it, and the period it covers, if known. Your own words come first: if you say "вот продажи по Арману", that sentence is stored as you said it and the assistant never overwrites it with its own guess. And the assistant cannot pass its own guess off as yours: words are marked as yours only if they are in what you actually wrote in that chat. Asking "what did our Yandex Direct ads bring?" does not tell Plank you use Yandex Direct.

A link is stored without its tracking part — everything after ? and # is dropped, and a Google document is kept by its id alone.

Logins and passwords are never stored in this list. If a message, a note or a description looks like a login or password, it is refused. A site where a login is saved shows only «вход сохранён». The login itself lives in «Доступы»: saving a login there marks its sites here — a shared login always, a personal one only in a workspace nobody else is in — and deleting it removes the mark. If the list noted that a login was written in plain text in a file, the assistant tells you where once it is saved, so you can remove that copy.

Why the assistant sometimes asks a question

Sometimes you'll get one short question after an answer:

• "Whom is this table about?" — when you add a source that looks like one already on the list (two sales sheets, say), and nothing says which client each is for. The answer is what lets the assistant pick the right one next time.
• "When should I use this?" — once per new source, so it ranks first for the right requests.

It asks at most one question at a time, never twice about the same source, and never in an automated run (a schedule, a webhook) — nobody is there to answer. If several sources fit a request and it cannot tell which you mean, it asks you which one instead of picking.

Seeing the list: the Sources page

Open Sources from the sidebar (on the phone: More → Sources). It used to be called Integrations; an old link or bookmark still opens it. It has two halves:

• Connections — the places the assistant logs in to by itself: connected services, and websites whose login is saved. Each shows whether its login works (Login works, Needs reconnecting, Connection not finished), or Login saved in Credentials. Nothing here is marked as working unless Plank actually knows it is. See Connecting your tools.
• Data — the files, reports, links and tables, the ones the assistant used most recently first. Filter by who added them (All, Sent by you, Made by the agent), by Topics, or search («revenue by doctor», «price list») — the search looks at the name, the description, when to use it and the topic. A source with No description is one the assistant will ask about next time you are in a chat. A data table shows Data table and how many rows it holds right now (counted when the page opens); the assistant does not ask about tables, so give one a description yourself if its name says too little. A table that was deleted disappears from the list, and what you wrote about it comes back if a table with the same name is created again.

Between the two halves, two rows open the pages that used to sit in the sidebar (until 12 October 2026):

• Credentials — your saved logins, passwords and keys. The row shows only how many are saved, never a value. See Credentials.
• Tables — shown once the workspace has a data table; it opens the list of tables. A single data table in Data also has an Open table button that shows its rows. See App data & databases.

Add source (the + on the phone) adds a Link or Google doc or a Website — paste the address and, if you like, say what it is and when the agent should look there. A file does not need adding: send it in a chat or upload it to Files, and it appears here on its own. A website's login goes in Credentials, never here.

When a connection stops working

A connection says No access since a date only when something that actually tried it failed: the check of a service's scripts, an MCP server's own status, Check on a saved login in Credentials, the workspace's shared browser being refused by a website, or Instagram access that was revoked or expired. Anything Plank has not tried stays a plain Connected — it is never shown as broken or as working on a guess. A connection that has answered says Working (an MCP server or a bot) or Login works.

Open the connection to see why, in one sentence (for example «The saved login failed its check — the key may be revoked or the password changed.»). Open in Credentials fixes a saved login; anything else gets Sign in again, which opens a chat with the assistant to help. Check connection asks again right away: it checks a saved login or re-reads a service's scripts, and for a source it has no way to check it says so — the status then updates the next time the agent uses it. A login someone saved for themselves is checked for them only: in a workspace with other members, its result is not shown as the source's status. Members who can only view do not see Check connection.

A Google document or Google service whose access was withdrawn on Google's side is not shown as No access since yet — Plank does not see that refusal; the assistant will tell you in the chat when it cannot open it.

A bot that receives messages (a webhook automation, for example a Telegram bot) is listed under Connections as Bot · receives messages, under the automation's name — rename the automation and the source follows, unless you gave the source a name of your own. A bot with no name is listed as Bot. A clinic's support-inbox intake is the platform's own and is not listed. Deleting the automation removes the bot from the list.

When the agent used it, and what else is connected

Open any source to see When the agent used it: each time a chat read it, made a result from it, or a schedule used it, with what was asked and the chat it happened in. A use in your own chat or in a shared chat links to that chat. A use in another member's private chat says only in a private chat — no title, no question and no link, so nobody learns what someone else asked. A schedule's use names the schedule and says automatically; so does a bot's — it names the bot and opens the bot's own page. The last 50 uses from the past six months are kept; uses from before this list started recording are not shown.

Also connected lists the scripts the source works through, other bots on the same script, and the schedules and skills whose text mentions it — open any of them from there. Schedules and skills are found by a mention in their text, so one that refers to the source in other words will not appear.

The agent needs your help

At the top of the page the assistant lists what it could not work out alone: a connection it can no longer get into (first), a question it asked in a chat that nobody answered, a link it found in a chat (Add or Not needed), a file you sent with no description, a source that needs a topic. On a computer the first three questions are on the page (and N more); on the phone, The agent needs your help opens them one question at a time (Save and next). Later puts a question away for a week — the assistant may still ask it in a chat.

The number next to Sources in the sidebar counts only what needs you now — a connection with no access and questions the assistant asked and is waiting on — never every undescribed file. Members who can only view see neither the questions nor the number.

Correcting the list

Open a source on the Sources page. Under What the agent knows about it are three lines — What it is, When the agent should look here and Who or what it is about — each with Edit. What you Save is yours: it is marked Written by you (or Written by a member, with the date), the assistant never rewrites it, and it reads it at the start of every chat. Clear a line to let the assistant suggest one again. A line the assistant wrote is marked Suggested by the agent · you can correct it, and the line under it says which chat it came from.

At the bottom of a source:

• Mark as outdated — it stays in the list, tagged outdated, but the assistant no longer offers it on its own. Current again undoes it.
• Remove from sources — the source leaves the list and the assistant stops offering it. The file, the login and the link stay where they are — nothing is deleted. Undo right after brings it back, and so does adding the same link again. A new upload of that file does not bring it back by itself.

Never type a password or key there — the page refuses it and sends you to Credentials. A source with a saved login shows the login's name only, says what is kept — a password, a key or a token — and has one Open in Credentials button; the value itself is never shown on this page. Members who can only view see the descriptions but cannot change them.

Or just say it in a chat: "this sheet is about Bekzat, not Arman", "that report is outdated", "use the new price list from now on". The assistant updates the source — your words replace the old ones.

Workspaces created before this list existed

The list was filled from what was already there — recent uploads, recent reports, connected integrations and websites. A file you attached in a chat also keeps what you wrote when you sent it ("here are the service reports"), so the assistant can find it in a new chat even when the file is just called report (12).xlsx. In each such workspace the assistant also does a one-time pass over your notes (AGENTS.md, index.md, knowledge/) and adds the sources it finds there. It does not edit those files. If a note has a login written in plain text, the assistant records only where it is written, not the login itself.

For the assistant — the tools

• plank_sources_find — search the list (by words, subject, kind, period). Every match comes back with its subject. When several match and the request does not say which, ask the person which one — never pick. A miss is not proof the data does not exist: look in the workspace too.
• plank_source_add — register a source the platform did not see arrive. Uploads, pasted links, reports, websites and integrations register themselves. Register only what you found exists (in notes, files, a link or login the person gave) — never a site or tool merely named in a question.
• plank_source_update — save what the person told you (subject, use_when, with from_person: true), or mark a source superseded / gone.
• from_person: true is checked: the words must be the person's own, copied as they wrote them in this chat. If they are not in the person's messages, the flag is dropped and the words are saved as yours — the tool result says so.
• When a tool result says ASK, ask that one question in your reply, in the person's language, and save the answer with plank_source_update.
• Never put a login, password, token or key into any field — it is refused. If notes contain one, set login_seen_in to the file it is in and copy nothing.