CommuteScout

Live road conditions across 37 states: a map, a route planner, and an AI assistant.
Also an MCP server, so your assistant can use it too.

Open the app · Add to Claude · Coverage · Self-hosting · Data · MCP tools · Architecture

No install, no account: commutescout.com

CommuteScout reads 53 official agency feeds (CHP dispatch, state DOT closures and incidents, chain controls, cameras, message signs, wildfire perimeters, road weather, toll prices) and turns them into one live picture of the road. Look at the map, plan a route and see what is actually on it, or ask about a drive in plain English. The same data is served over MCP, so Claude and other AI assistants can use it as a tool instead of guessing about traffic.

What you get

• A live national map: incidents by type, closures by class, chain controls, wildfires with real burn footprints, roadside weather stations, roughly 18,000 traffic cameras, and every message sign currently displaying something.
• Toll and express-lane pricing: current rates on tolled corridors and bridges, drawn along the actual carriageway with hand-verified gantry positions, so a price tag never floats over the wrong road.
• A route planner that knows the roads: autocomplete, route options, turn-by-turn directions, live conditions along the way, and print, GPX, KML, or share-link export.
• An assistant that reads the feeds: plan a route, tap a suggested question, and the answer streams in from the same live data with per-source timestamps.
• Plain-English incident detail: CHP dispatch logs are translated from radio shorthand into readable timelines, with each unit's arrival and clearance in order.
• Watch areas: draw a circle, polygon, or route corridor and get a push or email alert when an incident, closure, chain control, or wildfire appears inside it.
• An MCP server: ten tools over curated corridors and regions, with a closure taxonomy that keeps a closed on-ramp from reading as a closed highway.
• Public evals: 91 golden questions on recorded fixtures, scored by an LLM judge that is never one of the evaluated models. The scorecard and its full history are committed to this repo.

Plan a trip. Autocomplete, route options, directions, print or export.

Ask about it. One tap on a suggested question; the answer reads the live feeds.

Or just look. Every layer toggleable, from full closures to blank signs.

Coverage

The map covers 37 states. Coverage is not uniform, because it is built from what each agency actually publishes: some states offer every layer keylessly, some publish roadwork only, and a few offer nothing usable. The map says so directly, shading unsupported states and naming what is missing rather than showing an empty region.

California is the deepest. It is the only state with CHP dispatch logs, per-lane closure detail, chain-control levels, and CAL FIRE perimeters. The assistant and the MCP tools answer for every covered state, but a California question gets that richer detail, while elsewhere they answer from the normalized state DOT feeds.

Per-state matrix of what is live and why the gaps exist: docs/state-coverage.md. States not yet integrated, with the reason for each: docs/state-expansion-audit.md.

Get started

The fastest way to use CommuteScout is the hosted app: commutescout.com. Nothing to run, always on the latest release, feeds already warm.

Add to Claude

Give Claude live road data with a custom connector:

https://mcp.commutescout.com/mcp

See it on the site: commutescout.com/mcp. Local stdio setup and the full tool reference: docs/mcp.md.

Self-hosting (advanced)

Everything here is MIT licensed and the core runs with zero accounts or keys:

{
  "mcpServers": {
    "commutescout": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/nicglazkov/commutescout", "ca-roads-mcp"]
    }
  }
}

The web app is pip install ".[demo]" then ca-roads-demo with an ANTHROPIC_API_KEY in the environment. For your own Cloud Run copy (small enough for the free tier most months), optional feed keys, map snapshot publishing, and the watch-areas setup, see docs/deploy.md.

Self-hosted deployments are supported on a best-effort basis: issues and PRs are very welcome, but there is no support guarantee for deployments I do not run.

The data

CHP incidents and dispatch logs; Caltrans and 30-plus other state DOT closures, incidents, cameras, message signs, and road weather; chain controls from California and the Pacific Northwest; WFIGS and CAL FIRE wildfires with perimeters; NWS alerts; USGS quakes; toll and express lane pricing; and optional TomTom and 511 SF Bay feeds.

Every response carries per-source data_as_of timestamps, and a failing feed is never silent: the last good data is served, flagged stale, with the error attached and surfaced all the way to the UI.

Full source table, refresh rates, and the closure taxonomy: docs/data-sources.md.

How good are the answers?

An eval suite scores the assistant against recorded fixtures: four scenarios (a Sierra storm day, a fire-closure day, a quiet day, and a byte-for-byte capture of a real fire-season day), 91 golden questions with ground truth including traps, and an LLM judge that is never an evaluated model.

Runs are triggered manually rather than on every release. Firing a full suite on each release turned out to cost more per month than the hosted assistant serves, so it now runs when a prompt or tool change actually warrants re-scoring. Every run appends to a committed history file, so the trend stays public: EVALS.md.

Under the hood

Three cleanly layered Python packages sharing one data spine: a feed layer with stale-while-revalidate caches and parsers that salvage complete records from truncated feeds, the MCP surface, and the web app.

The map does not boot through the API. A publisher builds the whole coverage area once per cycle and uploads pre-gzipped snapshots to object storage behind a CDN, so first paint is an edge-cached static file and no visitor request waits on a server assembling JSON. A map left open on a wall monitor keeps updating in place indefinitely.

Diagram and design notes: docs/architecture.md.

Contributing

PRs welcome. The test suite is fixture-based and runs without network access. Start with CONTRIBUTING.md, and see adding a data source if you want to wire up a new feed.

License & sustainability

CommuteScout is MIT licensed: the map, the planner, the MCP server, and every data parser, with no open-core carve-outs. The hosted app at commutescout.com will soon offer optional premium features (deeper history, more alerts); that is what funds the servers and keeps the free tier free.

Disclaimer

Data comes from CHP, Caltrans and the other state DOTs listed in docs/state-coverage.md, plus WFIGS, CAL FIRE, NWS, and USGS. Not affiliated with any agency. Conditions change faster than any feed; verify before you drive (511 or your state DOT, and quickmap.dot.ca.gov in California).

Map tiles, routing, and place-name lookup come from Stadia Maps (data (c) OpenStreetMap contributors), so that service sees the coordinates involved. Fonts and map libraries are served locally.