Connect Claude Code

Connect Claude Code to NavFlow so it can read your data over MCP. Three ways: the plugin (one install — MCP read-back plus session capture), stdio (the client spawns a local proxy — simplest for an agent on the same machine as the daemon), or HTTP (connect to a running MCP server — for a remote or deployed instance).

Choose a connection

Plugin (recommended)

One install wires the MCP server and streams your sessions into the claude_code source:


/plugin marketplace add glassflow/navflow
/plugin install navflow@navflow

The install prompts for the NavFlow URL (the default matches a local navflow up), an optional auth token, and a capture toggle. Requires the navflow package on PATH (uv tool install navflow) — the plugin spawns navflow-mcp like the stdio path does.

stdio (local)

The client runs navflow-mcp, which proxies to the daemon. The navflow package must be installed where Claude Code runs.


claude mcp add navflow \
  --env NAVFLOWD_URL=http://localhost:8787 \
  -- navflow-mcp

If the daemon requires a token, add --env NAVFLOW_AUTH_TOKEN=<token>.

HTTP (remote)

First run the MCP server (or use a deployed instance that already serves /mcp):


navflow mcp --transport streamable-http --port 8788 --navflowd http://localhost:8787

Then add it by URL:


claude mcp add --transport http navflow http://localhost:8788/mcp

If the instance sets NAVFLOW_AUTH_TOKEN, pass it:


claude mcp add --transport http navflow https://navflow.example.com/mcp \
  --header "Authorization: Bearer <token>"

The console’s Agents → Connect tab generates these commands for your instance, with the endpoint and token filled in.

Verify the connection


claude mcp list
# navflow: connected

Use it

In a Claude Code session, refer to NavFlow by name. The agent discovers and calls the tools:

Use navflow: which sources am I ingesting, and what does each contain?

Use navflow to describe source:metrics and tell me the entities it carries.

With navflow, create a view correlating my logs and metrics by service, then show me the timeline for service=checkout.

The agent uses catalog_list / catalog_describe to learn the data, derive to compose a view, query to read it, and remember to write back observations.

Watch the calls

In the NavFlow console, Agents → Reads lists each read as it happens, tagged client = mcp.

Other clients

The same endpoint works with any MCP client (Codex, Claude Desktop, Cursor). See Connecting agents for the config formats.