map-mcp

Drive and perceive an existing, live MapLibre GL map from an AI agent (over MCP) or a human CLI — query the rendered features, read the viewport, click and read popups, navigate, toggle layers. The agent and the CLI act on the same map a person is looking at, with parity by construction.

It does not generate maps. Other geo-MCP servers (gis-mcp, Mapbox, CARTO) create maps or call GIS operations; map-mcp reaches into a map that's already on screen. Think of it as the agent-native counterpart to MapGrab: same live-map access, but conversational over MCP instead of written as test code.

Status: v1, MapLibre GL only. Cooperation-required (your app adds a one-line hook). A no-cooperation path and other map libraries are future work.

A LangChain agent finds the most-populous visible city and navigates to it — perceiving and driving the live map through map-mcp.

Install

uvx map-mcp --help          # or: pip install map-mcp

Related MCP server: QGIS MCP

Quickstart

1. Start the bridge + MCP server. It prints a WebSocket URL and a per-session token.

   map-mcp serve

2. Add the hook to your MapLibre page (map is your existing maplibregl.Map):

   <script src="map-mcp-hook.js"></script>
   <script>
     mapMcp.register(map, { url: "ws://127.0.0.1:8765", token: "PASTE_TOKEN" });
   </script>

3. Point your agent at the MCP server (stdio by default; --transport http for HTTP/SSE). Or drive it yourself from the terminal:

   map-mcp call get_viewport
   map-mcp call query_rendered_features --params '{"point":[12.5,41.9]}'

There's a runnable sample app in examples/sample_app/.

Example: a LangChain agent

examples/langchain_agent.py is a real agent — a langgraph ReAct agent on any tool-capable model — that drives the live map through map-mcp's operations (it's how the GIF above was made). Give it the map-mcp tools and a task, and it perceives and navigates the map itself:

export OPENROUTER_API_KEY=sk-or-...
uvx playwright install chromium          # one-time, for the live browser
uv run --extra demo python examples/langchain_agent.py

It defaults to an OpenRouter model (set OPENROUTER_MODEL to change). The agent's tools are thin wrappers over the same CoreOps the MCP server exposes — so anything the agent does, the CLI does too.

Tools (the operation surface)

The agent's MCP tools and the CLI's call operations are exactly the same set:

Perception returns structured feature properties (GeoJSON-shaped) — agents reason over properties, not pixels. screenshot is optional.

* needs the map created with preserveDrawingBuffer: true (see hook/snippet.md).

How it works

The hook connects out to a loopback WebSocket the map-mcp process runs. The MCP tools and the CLI are thin frontends over one shared core-operations layer, so any operation one can do, the other can too.

Security model (local-only). The bridge binds 127.0.0.1 only, so nothing off your machine can reach it. Browsers do not apply same-origin policy to WebSocket connections, so the per-session token is the security boundary: only a page that presents it can drive your map. Treat the token like a secret — the convenience ?token= pattern in the example leaks it via browser history and server logs, so for anything sensitive paste the token into the page rather than the URL. A hardened Origin allowlist is future work.

Scope (v1)

• In: MapLibre GL; the operations above; stdio + HTTP/SSE; a human CLI with parity.

• Out: generating maps; a hosted service; non-map visualizations; other map libraries (Leaflet/deck.gl) and a no-cooperation (Playwright) path are future work.

License

MIT