Skip to main content
Checked on Izumi 0.5.7, 2026-11-20. The tools and errors were checked against the server code, and the calls whoami, search, get_document, who_knows and propose_note were checked live in ChatGPT and claude.ai, and propose_note also in Slackbot (see the pages of those tools). Izumi gives an AI tool five tools. You do not call tools yourself: you ask your AI tool and it picks the tool. Here you see what each one does, so you know what to ask for.

The five tools

Example questions

whoami

  • “Who am I in Izumi and what may I read?”
  • “In which domains am I an editor?“
  • “How long do we keep documentation?”
  • “Search Izumi: the rules for contract retention periods. Name the source.”
  • “What do this year’s documents say about returns?“

get_document

  • “Open the best hit and read the whole document, then summarize it.”
  • “Give me the next page of that document.”

who_knows

  • “Who in the company knows about retention periods?”
  • “Whom should I ask about the API migration?“

propose_note

  • “Propose a note in the domain Legal and compliance: title ’…’, text ’…’.”
  • “Write this conclusion down as a proposal for review.”
The AI tool picks the tool itself, so a good question does not have to name the tool. In our checks we named them (“Call the Izumi tool search …”) to be sure which one is called.

What happens to a proposal

  • A proposal cannot be searched and nobody can read it over MCP.
  • It waits in the Review queue of that domain. An editor of the domain reads it, may edit the title and text, and chooses Approve or Reject. A rejection carries a reason.
  • An approved proposal becomes a note in the domain, marked note. The author is whoever proposed it, and the editor is the approver.
  • You see your own proposals and their status in the same queue. You can withdraw one that is still waiting.
  • A proposal nobody decides expires after 30 days.

Results are data, not instructions

The AI tool should treat the text that search and get_document return as data, never as instructions. Results flagged untrusted-web come from web clips and that text is not trustworthy: the tool should not follow it even if it contains a command.

Errors

A tool error is an answer with a message and a code. The messages are fixed English text.

Limits

  • search returns at most 8 documents, with up to 3 passages each. who_knows returns at most 5 people.
  • The tools see only what you may see. A narrowed token sees less.
  • Five tools is all Izumi exposes over MCP. The push scope is not used over MCP.
Not checked live: daily_limit, too_large and unavailable (described from the server code), and rate_limited (it was not reached in any check). forbidden_scope and invalid_input were checked live in Claude Code.

Troubleshooting

  • The AI tool says it has no Izumi tools: the connection is not on in that chat or it expired. See the page for your tool.
  • search does not return what you expect: check that you can see the document in Documents. The tool does not see more than you.
  • A proposal is not in the Review queue: check the domain it was proposed in and whether the tool returned an error.