> ## Documentation Index
> Fetch the complete documentation index at: https://docs.anarlog.so/llms.txt
> Use this file to discover all available pages before exploring further.

# MCP for agents

> Connect an MCP client to Anarlog meeting context over hosted OAuth or local stdio.

The Anarlog plugin prefers local data when your agent can access it. Use local MCP or the `anarlog` CLI on your computer. Remote agents connect to hosted Cloud MCP at `https://api.anarlog.so/mcp` with OAuth.

Local access needs no Cloud login, Pro subscription, or completed sync. Cloud only has opted-in readable snapshots, uploaded separately from encrypted Cloud Sync.

## Configure the plugin

Install the **Anarlog** plugin and [local CLI](/installation) for local access. For Cloud access, enable **Cloud API & Connectors** and use the host's connection control to approve OAuth. Some MCP hosts prompt on first cloud tool use. See [Cloud API and connectors](/reference/api-cloud#remote-mcp).

Clients that do not install plugins can add:

```json theme={null}
{
  "mcpServers": {
    "anarlog": {
      "url": "https://api.anarlog.so/mcp"
    }
  }
}
```

## Optional local MCP

The local server communicates over standard input and output. It opens no port and sends no meeting data to Anarlog's servers.

Open the desktop app once so it can create the database. The app does not need to stay open after that.

```json theme={null}
{
  "mcpServers": {
    "anarlog": {
      "command": "anarlog",
      "args": ["mcp"]
    }
  }
}
```

If `anarlog` is not on the client's `PATH`, use the executable's absolute path as `command`. Restart the client after changing its configuration.

For a custom database path:

```json theme={null}
{
  "mcpServers": {
    "anarlog": {
      "command": "anarlog",
      "args": ["--db-path", "/path/to/app.db", "mcp"]
    }
  }
}
```

<Note>
  If you configure both local and Cloud MCP, give them distinct server names so the agent can identify the source. Prefer local data and keep each meeting's subsequent reads on the selected source.
</Note>

## Use tools deliberately

Start with `list_meetings`, then pass a returned `id` to `get_meeting`. Call `get_meeting_transcript` only when the notes and summaries do not answer the question. Report only meetings the tools return and identify the source. An empty search only establishes that no meetings matched in that source. Check an available alternate source for missing meetings unless the user restricted access; do not invent titles or silently merge different versions.

Report sync status or snapshot freshness only when a tool provides it. A meeting's `updated_at` is not a last-sync timestamp. Local database errors need to be reported rather than hidden by a Cloud fallback.

Transcript pages default to 200 words and cap at 500. Follow `pagination.next_offset` only when you need more context.

Use `get_recurring_meeting_history` with a known meeting ID when a task depends on earlier meetings in the same series. Use `export_meeting` only when the task explicitly needs the complete record including transcripts.

Cloud MCP is read-only. To persist an edit, use the local CLI or local MCP `propose_summary_edit` or `propose_memo_edit`. The tool returns a pending proposal. Do not claim the meeting changed. Use `list_proposals` and `get_proposal` to inspect staged work. `decline_proposal` discards a pending proposal; applied or already-declined proposals cannot be declined.

The server also exposes MCP resources (`anarlog://meetings/...`) for Markdown or plain-text context when the client prefers resources over tools.

See the [MCP reference](/reference/mcp) for exact parameters and resources.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.