tesla-grokbot-connector

Read-only Tesla Fleet API MCP for Grok Bot. Each person runs this on their own Grok Bot computer. The Bot starts it as a stdio command. There is no public host and no marketplace tile.

Compatible with Tesla. This project is not affiliated with, endorsed by, or sponsored by Tesla, Inc. Do not name your developer app anything that implies an official Tesla product (use something like “GrokBot Vehicle Bridge”).

Install and attach: follow INSTALL.md. Put the client id, client secret, and token key on the computer, run tesla-mcp auth with Take over, then add a custom stdio command in chat.

What it does

Ask how charged the car is, how far it drove this week, or whether range is dropping. The server can:

• list vehicles (VIN masked)

• read battery, charging, climate, odometer, and software when the car is already online

• return the last local snapshot when the car is asleep, without waking it

• keep that history in SQLite and answer weekly mileage, range trend, and a low-charge check from it

• wake the car only through tesla_wake_and_refresh, and only inside a small budget

• optionally poll every 1–2 hours, and only if the car is already online

Related MCP server: mcp-teslamate-fleet

What it does not do

• The tesla-mcp server does not lock, unlock, honk, start climate, or start and stop charging. Those tools are not registered on it. The only Fleet API write on this server is wake_up, and only from tesla_wake_and_refresh.

• It does not use a virtual key or tesla-http-proxy.

• It does not stream Fleet Telemetry and it does not talk to Tesla Energy products.

• It does not request vehicle_cmds or user_data.

• It does not wake the car for a status question, a report, or tesla-mcp poll.

Optional vehicle commands

Lock, climate, charge, lights, horn, and sentry on are a separate program, tesla-commands-mcp. You add it as its own connector, tesla-commands, beside tesla. It is optional and off by default. Installing or running tesla-mcp does not add those tools and does not request vehicle_cmds.

Commands stay in dry-run until you set TESLA_COMMANDS_DRY_RUN=false (0, no, and off also turn it off). A blank or unrecognized value stays in dry-run. An empty allowlist allows none of them. Read docs/COMMANDS.md before you turn one on. That page covers the risks, the scope, pairing the virtual key, the signing proxy, and how to remove the key from the car.

Charging history is off. Tesla only exposes it through vehicle_charging_cmds, which also allows charge control. Turn it on only if you accept that grant. This server still will not call charge control.

How this compares to other Tesla MCP servers

Several open-source Tesla MCP servers already exist. The notes below are from their READMEs on 22 Oct 2026.

• scald/tesla-mcp is a small TypeScript MCP server for Claude Desktop. Its tools wake the car and return basic vehicle info (wake_up, refresh_vehicles, debug_vehicles).

• ysrdevs/tesla-mcp is Python. The README describes 96 tools covering the Fleet API, including lock/unlock, climate, and charging commands.

• MyTeslaMate/mcp-tesla is mainly a hosted service: you sign in through https://mcp.myteslamate.com/mcp. The README describes full vehicle-command access plus TeslaMate tools (98 Fleet API tools and 9 TeslaMate tools). Self-hosting is documented as an advanced option.

This project is built for Grok Bot. The tesla-mcp connector is strictly read-only: it has no command tools, no virtual key, and no tesla-http-proxy. An optional commands program lives beside it and is off by default (docs/COMMANDS.md). The read-only connector runs locally on the Grok Bot computer as a stdio process, with no hosted server. Wakes go through one tool with a local budget, so Fleet API wake charges stay capped. Mileage, range, and low-charge answers come from a SQLite history on that same computer.

Cost and billing

Prices from the Fleet API pricing page, checked 22 Oct 2026:

Authentication calls are not billed. Responses under HTTP 500 are billed, including 408 when the car is asleep. There is no sandbox. A live call hits your car and your bill.

A 2-hour poll is about 360 state checks a month, plus about 180 data reads if the car is online half the time ($1.08). About 10 questions a day add $0.60. One wake a day adds $0.60. That is about $2.30 / month, inside the $10 monthly discount, so the net is often $0. Which category GET /vehicles/{vin} and charging history use is per-endpoint in Tesla's docs and is treated as data here until you confirm it.

Set a billing limit around $10 and add a card. The dashboard default is $0, and an app with no payment method is disabled. Alerts arrive at 80% and 100%. Crossing the limit suspends the API. Payment works in the US and most of Europe, including IE and GB.

Wakes are also capped by Tesla at 3 per minute. A loop at that cap is about $86 a day. This server defaults to 3 wakes a day and 1 wake per 30 minutes, and it will not let you configure more than 3 per minute.

tesla_usage shows the local estimate. It is not Tesla's invoice.

Tools

vin is optional when the account has one car. Pass a display name or the masked VIN when there are several. Distances from the Fleet API are miles. Temperatures are Celsius. Every tool is marked read-only. tesla_wake_and_refresh is also marked non-idempotent because it costs money and wakes the car.

tesla_get_charging_history accepts a date (YYYY-MM-DD) or an ISO 8601 time for start and end. Leave both out for the last 365 days. A date-only start is midnight UTC, and a date-only end is 23:59:59 UTC. Tesla rejects a single request longer than one year (Date range cannot exceed 1 year), so a longer range is loaded in consecutive windows of at most 365 days, up to 10 years. Duplicate sessions are dropped and the newest session is returned first. The start and end in the reply are the UTC times that were queried. When Tesla rejects a request, the reply includes Tesla's explanation rather than only the HTTP status.

Scopes

Granted scopes are read from the access token at startup. A tool whose scope was not granted is not registered. Run tesla-mcp scopes to see the diff. Changing scopes needs tesla-mcp auth again (prompt_missing_scopes is set).

Poll and routines

tesla-mcp poll lists vehicles and, only when state is online, reads vehicle_data. Asleep and offline cars get a state-only row. It never calls wake_up.

Ask the Bot to run this every 1–2 hours:

uv run --directory /workspace/tesla-grokbot-connector tesla-mcp poll

tesla-mcp poll --loop stays in the foreground. Intervals under 1 hour are raised to 1 hour. A routine that runs the one-shot command is the better fit.

Example routines, all of them using the last snapshot rather than a wake:

• Weekly mileage (Monday 08:00): tesla_report with weekly_mileage. Gaps while the car slept are split evenly between readings.

• Range trend: tesla_report with range_trend. Rated range is scaled to 100% charge (range / SoC) over 30 and 90 days.

• Low charge (21:00): tesla_report with low_charge. Alerts under 30% when the car is unplugged. If the snapshot is older than 6 hours, the report says so.

• Charging summary (1st of the month): only after the charging-history opt-in. Compares the previous month with the one before it. If Tesla rejects the request, the reply includes Tesla's explanation.

Sign-in and token file

Client id and secret come from a secret-request or from config.env. The AES key comes from TESLA_TOKEN_KEY and is not written into the token file. Refresh tokens are single-use. The new refresh token is saved atomically before the new access token is used. The previous refresh token stays valid for 24 hours. A routine at least once a week keeps the chain alive (the poll does this whenever it runs). 401 login_required means run tesla-mcp auth again.

Every Bot on your Cursor account shares this computer. Encryption keeps the tokens unreadable without the key. It does not isolate one Bot from another Bot that can read the same secret. That is acceptable here because the scopes are read-only. Say so once.

Regions: TESLA_REGION=na, eu, or cn. A wrong region returns HTTP 421. Details: docs/partner-registration.md.

Privacy

• Location is off unless you opt in. The default precision is about 1 km. Precise coordinates are not written to SQLite. Logs do not contain coordinates or tokens.

• Logs are tool name, masked VIN, status, and cost category.

• tesla-mcp logout deletes the token file. Add --delete-history to delete snapshots as well.

• Revoke the app at https://auth.tesla.com/user/revoke/consent?revoke_client_id=<id> or in the Tesla app under Security & Privacy → Third-Party Apps.

• The billing limit is a hard spend cap.

See SECURITY.md.

Development

uv sync
make test
make lint

Tests mock the Fleet API. They do not call Tesla. scripts/live_smoke.py refuses to run unless TESLA_LIVE_SMOKE=1, and it still does not wake the car. Expect a few data calls, well under $1.

Before you turn this on

1. Create the developer app, host the public key, and register the partner domain. docs/partner-registration.md.

2. Add a card and set a billing limit around $10.

3. Decide region (na, eu, or cn), whether location is in, and whether the charging-history scope is worth the broader grant.

4. Confirm localhost is accepted as a redirect URI. If it is not, use tesla-mcp auth --paste with the Pages callback.html.

Unconfirmed with Tesla, as of this version: github.io as a partner domain, the exact vehicle_data endpoints names, software-update field names on every firmware, whether home charging appears in charging history, and whether reading an already-awake car delays sleep. Virtual-key-free vehicle_data is what Tesla's proxy documents for that endpoint; confirm it on your car.

License

MIT. Copyright (c) 2026 Naoise Irwin. See LICENSE.