MCP overview
What Gurusup gives your AI assistant once you connect it, how the assistant is told to behave, and who can use it, so you know what to expect.
Gurusup does not have a chat window of its own. You use it through your AI assistant, such as Claude, ChatGPT, Cursor or Codex. Gurusup connects to the assistant over MCP, a standard way for an assistant to use outside tools. The assistant asks Gurusup, Gurusup returns evidence from your organization's memory, and the assistant writes the answer. Gurusup does not write answers.
What your assistant can do with Gurusup
Once connected, your assistant gets the tools listed in the Tools reference. In plain terms, it can do the following.
| Ability | What it means |
|---|---|
| Ask | The assistant asks a question. Gurusup returns excerpts, the facts behind them, and the title and reference of the pages they come from. This is the main way to use Gurusup. See Ask questions. |
| Search | The assistant looks for pages that match a topic and gets a short list of hits. |
| Retrieve | The assistant opens specific pages found by asking or searching, when it needs more of the text. It asks for a few at a time. |
| Save knowledge | The assistant can send Gurusup a new fact you told it, such as a decision or a policy. It is meant to do this only after you ask or confirm. |
| Use skills | The assistant sees the skills you have access to and loads one when it needs it. When you ask, it can also submit a skill you wrote. A submitted skill is not usable until an admin approves it. See Skills. |
| Ask a person | When Gurusup has no answer, the assistant can offer to ask a colleague who is set up as a contact. It shows you who first and sends the request after you confirm. If you ask to be contacted yourself, that request counts as the confirmation. See Contacts and My requests. |
| Operate Gurusup | The assistant can run things in Gurusup for you from the chat: connect a source, check and build your memory, publish skills, manage members and read your usage. Admins get the full set and members a smaller one. It needs the ?profile=catalog address (see Tools reference). Credentials never go through the chat: the assistant gives you a link to Gurusup's own page. See Operate Gurusup from your assistant. |
| Report a problem or send feedback | The assistant can report a failure it can point to, and suggest an improvement when nothing failed. Neither changes what Brain holds. |
Several things change something outside the conversation: saving knowledge, submitting a skill, asking a person, sending a report or feedback, and any operation tool that changes something, such as connecting a source. Reports and feedback only record a note for the Gurusup team, but they still leave the conversation. No tool edits or deletes knowledge that is already saved.
How the assistant is told to behave
Connecting a server gives the assistant tools. What it does with them comes from two places: Gurusup's own instructions, which it receives when it calls init, and the short instruction you save in your assistant (see Setup).
The instruction you save only says to use Gurusup for company questions and to call init first. The detailed rules arrive with init, so they always match the current version of Gurusup. In summary, the assistant is told to:
- Call
initonce per task or session, before the first read, and treat the result as the rules to follow. - Ask Gurusup first with
ask_company_memory, giving the subject of your question and the question itself. Use search and retrieve only as a fallback. - Answer only from what Gurusup returns, and cite the titles and links Gurusup returned.
- Keep scores, rankings and other internal signals out of the answer, unless you ask for search diagnostics.
- Say what Gurusup's evidence supports and what it does not, instead of guessing. Not seeing a fact in the excerpts does not prove Brain does not hold it, so the assistant says the returned evidence does not establish it.
- Say so when nothing comes back, instead of guessing.
- Not present a service, sign-in or connection error as an empty result. It should explain the problem plainly, without backend details.
- Offer to ask a colleague only when a company fact it needs is still unsupported after a focused search, and only a contact that Gurusup returns after a preview. If no contact is available, it says so and does not suggest a person.
- When
learnis available, to run something in Gurusup, read Gurusup's skill for the job with it, use the tool for that operation, and keep every key, token and approval on Gurusup's own page, never in the chat. - Report a failure once it is confirmed, and send feedback once at the end of a task.
The instruction is a request to the assistant, not a lock. A different model or client can follow it less closely. If an assistant does not use Gurusup, see Troubleshooting.
Your assistant only uses Gurusup if it has been told to. Adding the server is not enough. The prompt on the docs home page saves that instruction for you, and each client page shows where to paste it by hand. See Setup.
What costs usage
Asking is not metered. Questions, searches and lookups do not use up your plan's usage. Content that Gurusup reads does, and that includes knowledge an assistant saves. When usage runs out, Gurusup keeps answering and new content waits for the next period. Saved knowledge is new content, so it waits too. See Plan and usage.
If adding new content is paused for your organization, Gurusup refuses the note and the assistant gets an error. When Gurusup accepts a saved note, the assistant tells you "Listo." and that the information will be available in a few minutes. That confirmation means Gurusup accepted the note. It does not mean the note is searchable yet.
Knowledge saved through an assistant cannot be edited or removed from the app afterwards.
Who can use it
Any member of the organization can connect an assistant and use Gurusup through it. You do not need to be an admin. The Gurusup MCP page is in the sidebar for everyone.
Gurusup answers only once an admin has built the organization's memory. Until then, the tools tell the assistant that Gurusup is not ready. Members cannot build the memory. See Troubleshooting.
Each person signs in to Gurusup from their own assistant, and one connection is one person in one organization. What the assistant can see depends on who signed in, not on which assistant it is. That means the categories an admin granted that person, plus any restricted source they were picked for. Admins reach every category, but a restricted source only if they were picked. See What an agent can see.
How sign-in works
The first time your assistant connects, it may open a browser window so you can sign in to Gurusup. After that, the assistant uses that sign-in. You do not paste a password or a key into the assistant. If you belong to more than one organization, you choose the organization when you sign in.
Approval prompts in your client
Each Gurusup tool is marked as read-only or not, and as destructive or not. Your client can use those marks to decide when to ask you before running a tool. Whether it asks, and how, is the client's choice, so the same tool can prompt you in one client and not in another. The Tools reference shows the marks.
What you get back
Gurusup returns evidence, and the assistant decides how to word the answer. A citation names a page of Brain, not the original document. See Sources and citations.