MCP endpoint

Mimir speaks MCP over HTTP at /mcp. It is stateless: every request carries its own token, so a stock reverse proxy works without SSE configuration.

Two values

bash

MIMIR_URL=https://mimir.example.com
MIMIR_TOKEN=mimir_...        # mint one under Settings → API keys

Repository header

Send X-Mimir-Repo: github.com/owner/repo so the agent lands in the same project another agent created yesterday. Without it, pass repo or project as a tool argument. A server never sees a working directory, so the client is the only thing that can compute this.

Raw HTTP

MCP here is JSON-RPC 2.0 over a plain POST. A client library hides this, but seeing it once makes debugging a token or a header much quicker.

bash

curl -s "$MIMIR_URL/mcp" \
  -H "Authorization: Bearer $MIMIR_TOKEN" \
  -H "Content-Type: application/json" \
  -H "X-Mimir-Repo: github.com/owner/repo" \
  -d '{
    "jsonrpc": "2.0",
    "id": 1,
    "method": "tools/call",
    "params": {
      "name": "memory_search",
      "arguments": { "query": "token rotation", "path_prefix": "auth" }
    }
  }'

json

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "results": [
      {
        "path": "auth/token-rotation",
        "title": "Rotating the API token",
        "summary": "Rotate in the admin panel, then redeploy.",
        "score": 0.0328
      }
    ],
    "content": [{ "type": "text", "text": "[{\"path\":\"auth/token-rotation\",…}]" }],
    "isError": false
  }
}

initialize answers with protocol version 2025-06-18 and tools.listChanged: false; ping answers with an empty result.
tools/list returns each tool’s name, description and JSON schema. No token scope narrows the list; a grant narrows what a call may touch.
A tool result carries the payload twice: as fields on result for a script, and as content[0].text for a standard MCP client.
A failed call is a JSON-RPC error with a service code in error.data.code, and still writes one audit row. An expired token is HTTP 401 with {"jsonrpc":"2.0","id":null,"error":{…}}; a rate limit is 429 with Retry-After.

Tools

Tool	Purpose
memory_tree	List a project’s paths with their one-line titles.
memory_read	Read one memory at L0, L1 or L2.
memory_search	Hybrid search, optionally scoped to a path prefix.
memory_digest	A budgeted session-start digest of the project.
memory_write	Create or replace a memory.
memory_ingest	Import a batch of files as memories.
memory_delete	Soft-delete a memory.
project_list	List the projects you can reach.

What to tell the agent

Call memory_digest once at the start of a session, then memory_search before re-deriving something the project already knows.
Call memory_write when you learn something durable: a convention, a gotcha, a decision and its reason. Not a log of what you just did.
Keep paths stable. One path is one memory; writing it again replaces the body.

Per-agent configuration lives under Agents in the sidebar.