uber-mcp

An MCP server for Uber rides. It lets an AI assistant (Claude, ChatGPT, Cursor, ...) find places, compare every ride option with Uber's live prices, request a ride once you have said yes to the exact price, follow the live ride (driver, car, plate, PIN, arrival time), cancel it, and read your trip history.

Unofficial. Not affiliated with or endorsed by Uber Technologies, Inc. Uber has no public API for booking rides as a rider; this server uses the same endpoints Uber's own rider website uses (m.uber.com and riders.uber.com), signed in as you, with the session of a browser you signed in to yourself. Uber can change those endpoints at any time, and the server then stops working until it is updated. Use it with your own account, for your own rides; Uber's terms of use apply to you, and whether automated use of your account is acceptable to Uber is your risk.

Requesting a ride spends your money and sends a real car. The server never does that on one call: every request is a preview first and a single-use confirm token second, within price caps you set, and booking stays off until you set them (Booking safety).

Requirements

• An Uber rider account with a payment method, in a city where Uber's rider website (m.uber.com) offers rides.

• uv (it fetches Python 3.12 if needed), or Docker for a server.

• Google Chrome for signing in (or Chromium, see Signing in).

Quick start

1. Sign in once. A browser window opens on Uber's own sign-in page; sign in as usual (a code, Google, Apple or a QR code). It closes once you are signed in, and the session is saved for the server.

   uvx --from "uber-mcp[login] @ git+https://github.com/benjamin-di/uber-mcp-server" uber-mcp login

2. Add the server to your client. It runs over stdio, started by the client. As configured here it can look up places, prices, rides and trips but cannot book: booking stays off until step 3.

   • Claude Code:

     claude mcp add uber -- uvx --from git+https://github.com/benjamin-di/uber-mcp-server uber-mcp

   • Claude Desktop, Cursor and most other clients: add to the client's MCP configuration

     {
       "mcpServers": {
         "uber": {
           "command": "uvx",
           "args": ["--from", "git+https://github.com/benjamin-di/uber-mcp-server", "uber-mcp"]
         }
       }
     }

3. To let it request rides, set both caps in the server's environment, in your account's currency: the most one ride may cost and the most everything it books may cost in any 24 hours, for example claude mcp add uber -e UBER_MAX_FARE="USD 80" -e UBER_MAX_FARE_24H="USD 200" -- uvx ..., or "env": { "UBER_MAX_FARE": "USD 80", "UBER_MAX_FARE_24H": "USD 200" } beside "args" (Booking safety).

4. Ask. "How much is an Uber from home to the airport right now?", "Book the cheapest one", "Where is my driver?", "What did my last three rides cost?".

Pin a release with git+https://github.com/benjamin-di/uber-mcp-server@vX.Y.Z. The package is not published on PyPI.

Related MCP server: Beckn Mobility MCP Server

Tools

A pickup or drop-off can be a saved place's label (Home), a place id from uber_places (Google place ids such as ChIJ..., so ids from a Google Maps tool work too), lat,lng, or free text (Uber's top match, echoed back with alternatives: check it). Airports list their terminals; pass a terminal's lat,lng to be set down there.

A price is {"display": "...", "cents": 4250, "currency": "USD", "kind": "fixed"} with Uber's own display string:

• fixed: Uber's upfront price, held for a few minutes (quote_expires_at).

• range: a metered taxi. The meter decides the fare; cents is the range's high end and low_cents the low end.

• estimate: neither of the above.

Options marked reserve_only can only be booked ahead, in the Uber app. Caps and spend are {"cents": 8000, "currency": "USD", "display": "USD 80.00"}. Errors are {"error": {code, next_step, retryable}}, and the next_step text tells the assistant what to do (Troubleshooting).

Booking safety

Enforced in code (src/uber_mcp/booking.py, the only module that can send a ride request or cancellation) and pinned by tests/test_safety.py and tests/test_booking.py:

• Off unless you set both caps. Without UBER_MAX_FARE (one ride) and UBER_MAX_FARE_24H (everything booked through this server in any 24 hours), uber_request_ride answers booking_disabled before asking Uber anything. Both use the same currency, and the per-ride cap cannot be above the 24-hour one; one without the other refuses to start.

• Preview, then confirm. The first call only previews and returns a confirm_token. The token works once, for 10 minutes, and only with exactly the arguments of its preview. The tool descriptions and the server's instructions tell the assistant to show the preview and wait for the person's explicit yes; the tools are marked destructive, so clients that ask before such calls do.

• Never above the approved price. At confirm the server asks Uber for a fresh price for the same option. If it is higher, or in another currency, nothing is booked and a fresh preview comes back (price_rose) for the person to approve. A lower price books. For a metered taxi the approved price is the range's high end; the meter decides. For an estimate Uber has not committed to a price, so the final fare can differ from the approved one.

• Within the caps. The ride's price must fit the per-ride cap and, together with the rides this server booked in the last 24 hours, the 24-hour cap, in the caps' currency (over_cap, with the reason). Cancelled rides still count; rides booked in the Uber app do not.

• One ride at a time. A preview or a confirm while a ride is active answers trip_active.

• No blind retries. When a request was sent but its answer is lost, the result is outcome_unknown: the assistant reads uber_ride_status, and the server never repeats a request on its own.

• Freeze. uber-mcp admin freeze [REASON] stops all booking at once and drops every pending preview; uber-mcp admin unfreeze resumes within the caps. Cancelling still works while frozen.

• Audit log. Every confirm, booked or not (refused with the reason, started, failed, refused by Uber, succeeded; cancels too), is appended to audit.jsonl with the product, the price and the trip id, never the places. uber-mcp admin audit [-n N] shows the latest.

• Only booking. No endpoint for payment methods, tips, ratings, profiles, saved places, scheduled rides, Uber Eats or account settings exists in the code.

Other safety

• Only Uber. Requests go only to m.uber.com and riders.uber.com, and only with the GraphQL operations listed in src/uber_mcp/uber.py; a redirect is never followed.

• Gentle. One request at a time, about one a second, a daily request budget, and a pause of 5 minutes (30 when it repeats within two hours) whenever Uber pushes back.

• Your account only. The server acts for the one signed-in rider. Every read of the account checks that it is still the rider who signed in. Place names, driver names and trip descriptions are treated as data, not instructions.

Privacy

Everything stays on the computer that runs the server, in its state folder (~/.local/state/uber-mcp on Linux, ~/Library/Application Support/uber-mcp on macOS, %LOCALAPPDATA%\uber-mcp on Windows, /state in the container; UBER_STATE_DIR overrides it):

Nothing is sent anywhere except Uber. What a tool returns (places, prices, the driver's name, the plate, the PIN) goes to your assistant, like anything else you ask it. Logs (stderr) have one line per tool call with its name, outcome and duration, never its arguments, cookies or Uber's error text. The account read includes your email address and phone number (Uber's own query asks for them); the server drops them. uber-mcp status shows only the first letter of the account's name. uber-mcp logout deletes the session and the browser profile.

Signing in

uber-mcp login opens Uber's sign-in page in your installed Google Chrome, using a separate profile in the state folder. You do every step yourself; nothing is clicked or typed for you, and the window shows only Uber's pages (and Google's or Apple's, for their sign-in buttons). Once Uber shows its rides page, the window opens Uber's trips site once (it signs in there through Uber's own single sign-on), reads your account on both with the same cookies, saves the session and closes. A running server picks up the new session with its next call.

• No Chrome? Run uvx --from "uber-mcp[login] @ git+https://github.com/benjamin-di/uber-mcp-server" patchright install chromium, then uber-mcp login --browser chromium.

• Uber decides how long a session lasts. The server reads both of Uber's sites every 4 hours (UBER_KEEPALIVE_SECONDS), which lets Uber renew its short-lived cookies; uber_status shows when each sign-in cookie expires. When the session ends, tools return session_expired: sign in again.

• uber-mcp status shows whether the saved session is signed in, without asking Uber anything.

• Every sign-in raises the session's generation; a newer sign-in always wins over cookies another process is still writing back.

On a server: the login service

A server has no screen for the sign-in window. The login browser image (ghcr.io/benjamin-di/uber-mcp-login-browser) runs Uber's sign-in page in a Chromium next to the server and shows it to the person through noVNC in their own browser. Nothing on the sign-in is automated there either.

1. Run the login service beside the server on a private network, with the same UBER_LOGIN_TOKEN (32 or more characters) on both, and UBER_LOGIN_URL on the server (Running it as a server).

2. Start a sign-in on the server and wait:

   docker exec -it uber-mcp uber-mcp admin login            # --layout phone for a phone-sized window

   It prints a link, UBER_LOGIN_VIEW_URL/view/<id>?t=<token>, that works for this one sign-in only.

3. The person opens the link and signs in to Uber on the page it shows. The service checks the session on both of Uber's sites, hands it to the server once and closes the browser; the server seals it (UBER_SESSION_KEY) and saves it. admin login gives up after 20 minutes (--minutes).

The link must reach the login service's port 8080. The default, http://localhost:8766, suits the compose file below (the service on the server's port 8766, loopback only) opened through an SSH port forward (ssh -L 8766:127.0.0.1:8766 <server>); set UBER_LOGIN_VIEW_URL to your own HTTPS address if you publish the service through a reverse proxy instead. Its API needs the bearer; the view page and the remote-desktop stream need the per-sign-in token.

Without the login service, sign in with uber-mcp login on a computer with a screen and copy its session.json into the server's state folder; the server adopts it on its next call. A server with UBER_SESSION_KEY reads only session.bin: sign in with the same key set (UBER_SESSION_KEY=... UBER_STATE_DIR=... uber-mcp login) and copy that.

Running it as a server

For remote clients (claude.ai, ChatGPT and the Claude and ChatGPT apps reach connectors over HTTPS), run it over streamable HTTP. It refuses to start without a bearer token:

docker run -d --name uber-mcp -p 127.0.0.1:8000:8000 --read-only --tmpfs /tmp \
  -e MCP_BEARER_TOKEN="$(openssl rand -hex 32)" \
  -e UBER_SESSION_KEY="$(openssl rand -base64 32)" \
  -v uber-state:/state \
  ghcr.io/benjamin-di/uber-mcp-server:latest

Like this it cannot book; add -e UBER_MAX_FARE=... -e UBER_MAX_FARE_24H=... to let it (Booking safety). Keep the bearer and the session key: a new key cannot read a session sealed with the old one (session_invalid, sign in again). Or run MCP_BEARER_TOKEN=... uvx --from git+https://github.com/benjamin-di/uber-mcp-server uber-mcp serve --host 127.0.0.1 --port 8000. The MCP endpoint is /mcp; /health is open for health checks (the image's HEALTHCHECK runs uber-mcp health). Put it behind your own HTTPS reverse proxy to reach it from outside, and give clients the bearer (Authorization: Bearer <token>).

The server image (amd64 and arm64) runs as uid 1000 on a read-only root filesystem: it writes only to /state, which holds everything in Privacy, and to a /tmp tmpfs. With the login service, as a compose file:

services:
  uber-mcp:
    image: ghcr.io/benjamin-di/uber-mcp-server:latest
    read_only: true
    tmpfs: [/tmp]
    ports: ["127.0.0.1:8000:8000"]
    volumes: [uber-state:/state]
    environment:
      MCP_BEARER_TOKEN: ${MCP_BEARER_TOKEN}
      UBER_SESSION_KEY: ${UBER_SESSION_KEY}
      # To let it book: both caps, in your account's currency (Booking safety).
      # UBER_MAX_FARE: USD 80
      # UBER_MAX_FARE_24H: USD 200
      UBER_LOGIN_URL: http://uber-mcp-login:8080
      UBER_LOGIN_TOKEN: ${UBER_LOGIN_TOKEN}
  uber-mcp-login:
    image: ghcr.io/benjamin-di/uber-mcp-login-browser:latest
    ports: ["127.0.0.1:8766:8080"]
    environment:
      UBER_LOGIN_TOKEN: ${UBER_LOGIN_TOKEN}
volumes:
  uber-state:

The login image (amd64) runs as uid 1000, keeps nothing (each sign-in gets a throwaway browser profile that is deleted when it ends) and refuses to start without UBER_LOGIN_TOKEN. It can stay stopped between sign-ins.

Configuration

The server (uber-mcp):

The login service (ghcr.io/benjamin-di/uber-mcp-login-browser, python -m uber_login):

Troubleshooting

How it works

Uber has no public rider API for this, so the server speaks GraphQL to the endpoints Uber's rider website uses, over plain HTTP, with the cookies of a browser session you signed in to and the GraphQL documents Uber's own web apps send. docs/uber-web-api.md records what is known about those endpoints; docs/DESIGN.md explains the decisions. When Uber changes its web API, calls fail with schema_changed until the server is updated.

Development

See CONTRIBUTING.md. Tests never contact Uber.

Credits

Built on what crafter-station/uber-cli worked out about Uber's rider web API: scripts/sync-operations.py is a port of its script that reads the GraphQL documents from Uber's own web bundles, and the price and ride request variables follow its code (MIT; THIRD_PARTY_NOTICES). Design ideas also came from AriOliv/uber-mcp (MIT).

License

MIT.