What Walkie's MCP server can do
Walkie runs two Model Context Protocol servers. A local one reads your meeting database on your own machine, for agents like Claude Desktop, Claude Code, Codex CLI, Gemini CLI. A remote HTTPS endpoint at mcp.trywalkie.com serves hosted agents like ChatGPT from the meetings you have synced, with each agent approved by name in the browser. Both are read-only, both expose the same tools — search_meetings and list_meetings and get_meeting — and both are available from Walkie Walkie Work upward on macOS. For the step-by-step setup, see the MCP setup guide.
The two servers
Which one you use depends on where your agent runs, not on what you want it to do. Both answer the same questions.
| Local server | Remote endpoint | |
|---|---|---|
| For | Claude Desktop, Claude Code, Codex CLI, Gemini CLI — agents that can start a process on your Mac | ChatGPT and Grok — hosted agents, which cannot |
| Reads | Your meeting database on disk — every meeting, including ones you never synced | Only the meetings already synced to your Walkie account |
| Reaches the network | No. It is a local process talking over stdio | Yes. One HTTPS URL, with the agent authorised in the browser |
What the agent can do
Two tools, identical on both servers, so an agent sees one contract however it connected.
- search_meetings — Search every recorded meeting by keyword or phrase. Matches meeting titles, your own notes, action items, participants, and what was actually said in the transcript.
- list_meetings — List every meeting as metadata only — id, title, date, length and participants, with no notes or transcript. Page through the whole archive with limit and offset.
- get_meeting — Read one meeting in full: the AI notes, action items, participants, your notes, and the speaker-labeled transcript.
Tool reference
Moved here from the setup guide: somebody following a wizard does not need an argument table, and somebody writing against the server does. Generated from the same definitions the servers implement, so it cannot drift from them.
| Tool | Arguments | Returns |
|---|---|---|
search_meetings | query (string), limit (number, default 10, max 20) | Matching meetings as ids with a snippet, the date, length, source and participants — never the whole archive. |
list_meetings | limit (number, default 50, max 200), offset (number) | A page of meetings plus the total, so an agent knows when to stop. On the local server this needs Full archive access switched on; without it the tool refuses and says so. |
get_meeting | id (string, from search_meetings) | One meeting. Transcript lines are labeled with the names you assigned, then Walkie's guesses, then a channel ordinal — and capped at 60,000 characters so a long meeting cannot blow up the agent's context. |
What it will not do
The tools are query-shaped on purpose: an agent searches, then reads one meeting, the same two moves you would make.
- Enumeration is opt-in. By default there is no list-everything endpoint, so nothing can quietly walk the archive. Turning on full archive access in Walkie → Settings → MCP adds
list_meetingsand lets search page — your call, off until you make it. - No writes. Nothing an agent calls can record, delete, edit, share, or change a setting.
- Off until you switch it on. Both servers ship disabled, per machine and per account. Downgrading switches the local one back off rather than leaving it running.
- Capped payloads. Search returns at most 20 meetings, and a transcript is truncated rather than handed over whole.
Setting it up
The full procedure — configuration for each client, the remote endpoint for ChatGPT, and what to do when it will not connect — is in the MCP setup guide. The short version for the local server is three steps: switch it on in Settings → General → MCP server, point your client at /Applications/Walkie.app/Contents/MacOS/walkie --mcp-serve, then restart the client.
Things to know
| Availability | Walkie Walkie Work and above |
| Platform | macOS only |
| Transport | stdio for the local server; Streamable HTTP for the remote endpoint |
| Permissions | Read-only. No tool mutates anything. |
| Status | Experimental |
Frequently asked questions about what Walkie's MCP server can do
Does this send my meetings to Anthropic or to Walkie's servers?
The local server sends nothing — it runs on your machine and reads your local database. What your AI client then does with the text it receives is governed by that client's own policy: if you ask Claude about a meeting, the relevant passages go to Claude to answer it. The remote endpoint is different by design, and answers from the meetings you have already synced.
Is this different from asking Walkie about a meeting in the app?
Yes. Asking inside Walkie works on the meeting in front of you, in Walkie's own window. MCP turns the whole archive into something your own agent can reach from wherever you already work, so a question can span many meetings and the answer can land in the document you are writing.
Can the agent see meetings I recorded before turning MCP on?
Yes. The server reads your existing meeting database as it stands — there is no indexing step and nothing to re-process. Everything already recorded is searchable the first time an agent connects.
Why is it Walkie Work and above?
The MCP server is tied to the meeting archive, and the archive-kept-forever guarantee it depends on starts at Walkie Walkie Work.