A Model Context Protocol (MCP) server that provides access to your TeslaMate database, allowing AI assistants to query Tesla vehicle data and analytics.
개요
Connect your AI assistant to your TeslaMate data. This is a Model Context Protocol (MCP) server. It reads your TeslaMate PostgreSQL database. It gives MCP clients (Claude Desktop, Cursor, and others) 57 tools, 11 prompts, and interactive charts. · Tool reference · Configuration · Deployment - "Is my battery really degrading?" — the server estimates usable capacity in kWh from your charging sessions. - "How much range do I lose while parked?" — the server finds vampire drain between drives. - "Show the route of my longest drive." — the server returns the GPS track, with an interactive map on chart-capable clients. - "Compare my driving this month with last month." — one call returns both windows, per metric. - "Can I reach Ankara (450 km) at 5°C without charging?" — the server estimates the arrival battery % from your own drives at that temperature. - "Why isn't my car sleeping?" — the server reads TeslaMate's state log and lists the periods the car stayed awake while parked.
README
What you can ask
- “Is my battery really degrading?” — the server estimates usable capacity in kWh from your charging sessions.
- “How much range do I lose while parked?” — the server finds vampire drain between drives.
- “Show the route of my longest drive.” — the server returns the GPS track, with an interactive map on chart-capable clients.
- “Compare my driving this month with last month.” — one call returns both windows, per metric.
- “Can I reach Ankara (450 km) at 5°C without charging?” — the server estimates the arrival battery % from your own drives at that temperature.
- “Why isn’t my car sleeping?” — the server reads TeslaMate’s state log and lists the periods the car stayed awake while parked.
- “What did my car do on Saturday?” — one call returns the drives, charges, and parked time in order.
Highlights
- 47 SQL report tools. Battery capacity and degradation, sleep and state history, a timeline of each day, trips, range estimates, fast-charging speeds, charging costs (with estimates for missing ones) and fuel savings, elevation, vampire drain, geofences, driving patterns, routes, and search tools for drives and charging sessions.
- 8 interactive charts (MCP Apps). Charging curves (one session, or several overlaid), battery degradation, drive and trip maps with elevation profiles, a monthly or yearly activity dashboard, a map of visited places, and consumption against temperature. On clients without chart support, the same tools return plain rows.
- Guidance for the assistant. MCP instructions and 11 prompts map questions to tools, and
get_unit_preferenceslets the assistant answer in the units set in TeslaMate. - Typed inputs and outputs. Each report tool declares its parameters and result columns in a
.tomlfile. The server validates the declarations at startup. Parameters bind as SQL placeholders, never as string concatenation. - Safe custom SQL.
run_sqlruns yourSELECTinside a PostgreSQLREAD ONLYtransaction, with timeouts and a row cap. The transaction always rolls back. - Optional cost writes, with confirmation. One opt-in tool writes charging costs. A column-scoped database grant limits what it can touch. Clients with MCP elicitation show a confirmation dialog first.
- Local-time reports. Set
REPORT_TIMEZONEand daily, weekly, and monthly buckets follow your local midnight. - Two transports. Run
teslamate-mcp stdiofor local clients. Runteslamate-mcp httpfor remote clients, with bearer-token auth and a/healthprobe.
Requirements
- TeslaMate with its PostgreSQL database.
- Python 3.11+ for a local install, or Docker for a server install.
Quick start — local (stdio)
-
Install:
git clone https://github.com/cobanov/teslamate-mcp.git cd teslamate-mcp cp env.example .env # set DATABASE_URL uv sync -
Add the server to your MCP client. Example for Claude Desktop or Cursor:
{ "mcpServers": { "teslamate": { "command": "uv", "args": ["--directory", "/path/to/teslamate-mcp", "run", "teslamate-mcp", "stdio"] } } }
Quick start — Docker
cp env.example .env # set DATABASE_URL, and set AUTH_TOKEN for remote use
docker compose up -d
The MCP endpoint is http://localhost:8888/mcp. The health probe is http://localhost:8888/health. It runs a SELECT 1 against the database and returns 503 when the database is unreachable.
A multi-arch image (linux/amd64, linux/arm64) is published to GHCR on each release:
docker run --rm -e DATABASE_URL=... -p 8888:8888 ghcr.io/cobanov/teslamate-mcp:latest
Deploy on Unraid, behind Cloudflare
You can reach this server from claude.ai, Claude Desktop, and mobile — with OAuth login and with no open ports on your home network.
The Unraid + Cloudflare deployment guide shows the full path:
- Run the container on Unraid with the included Community Apps template.
- Publish it through a Cloudflare Tunnel. No inbound firewall rules are necessary.
- Put a Cloudflare Zero Trust MCP Server Portal in front. The portal handles OAuth for Claude clients and sends the bearer token upstream.
The guide includes the exact dashboard fields, verification commands for each phase, and a troubleshooting table.
The portal rewrites
ui://resource URIs, so the interactive charts (show_*tools) do not render through it. To get them, connect Claude through Cloudflare Access directly. The guide’s direct connection section shows how.
Tools
Each report tool accepts optional filters: car_name everywhere, plus days, limit, and thresholds where they apply. A call with no arguments returns the full report.
Tools return metric values (km, km/h, °C, bar, m) and rated range. get_unit_preferences reads the units chosen in TeslaMate’s settings (km or mi, °C or °F, bar or psi, rated or ideal range) so the assistant can convert when it answers; the charts stay metric.
The server also sends MCP instructions: how to read units, time zones and costs, and which tool answers which kind of question.
Reports (22)
| Group | Tools |
|---|---|
| Vehicle | get_basic_car_information, get_current_car_status, get_software_update_history, get_unit_preferences |
| Battery | get_battery_health_summary, get_battery_degradation_over_time, get_daily_battery_usage_patterns, get_tire_pressure_weekly_trends |
| Driving | get_monthly_driving_summary, get_daily_driving_patterns, get_longest_drives_by_distance, get_total_distance_and_efficiency, get_drive_summary_per_day, get_visited_places |
| Efficiency | get_efficiency_by_month_and_temperature, get_average_efficiency_by_temperature, get_efficiency_by_elevation, get_drive_efficiency_points, get_unusual_power_consumption |
| Charging | get_charging_by_location, get_all_charging_sessions_summary, get_most_visited_locations |
Insights (8)
| Tool | What it returns |
|---|---|
get_battery_capacity_trend |
Usable capacity in kWh, estimated from charging sessions (energy added ÷ SOC gained), per month and car |
get_vampire_drain |
Range lost while parked between drives; gaps that contain a charge do not count |
get_charging_efficiency |
kWh added vs kWh drawn per car, split into AC and DC |
get_charging_by_geofence |
Charging totals per TeslaMate geofence, plus an “Ungeofenced” group |
get_soc_hygiene |
Share of samples above 80% and below 20% state of charge |
get_period_comparison |
The last N days vs the N days before, one row per metric |
get_activity_report |
A month by day or a year by month: drives, distance, consumption, charging, and cost per bucket, empty days included |
get_trip_energy_estimate |
“Will I make it?”: energy and battery % a drive of a given distance takes, from this car’s own drives at a similar temperature (optionally motorway drives only), with a conservative figure and the arrival battery % |
Timeline and car state (3)
TeslaMate logs whether each car is online, asleep, or offline. get_current_car_status reads that log too, and reports whether the car is driving, charging, online, asleep, or offline right now.
| Tool | What it returns |
|---|---|
get_timeline |
What the car did, in order: drives, charges, the time parked between them, and software updates, with places, battery %, kWh, and cost; answers “what did my car do on Saturday?” |
get_state_history |
Per car and local day: hours online, asleep, and offline, driving and charging hours, idle awake hours (online while parked and not charging), and wake-ups |
get_idle_awake_periods |
The online periods that kept the car awake while parked, longest first, with where it was parked; answers “why isn’t my car sleeping?” |
Search and detail (5)
| Tool | What it returns |
|---|---|
search_drives |
Drives filtered by date range, location text, distance, and car; sortable |
search_charging_sessions |
Charging sessions filtered by date range, location, energy, and car |
get_drive_details |
Full statistics for one drive |
get_drive_route |
GPS track points for one drive, downsampled |
get_charging_curve |
Power and SOC curve for one charging session, downsampled |
Costs (3)
A session with no cost recorded counts as unknown, not free: averages per kWh use only the sessions that have a cost, and each report says how many sessions lack one.
| Tool | What it returns |
|---|---|
get_charging_costs |
Costs grouped by month, location, or car, with the number of sessions that have no cost |
get_charging_cost_estimates |
An estimated cost for each session with none: free Supercharging, a price you give, the geofence tariff, what you paid at the same place, or the typical price for that charger type |
get_fuel_savings |
What the same distance would have cost in fuel, against what charging cost, per car |
Trips (3)
TeslaMate ends a drive every time the car parks, so a road trip with a WC break and a charging stop is several drives. The trip tools merge a car’s consecutive drives while each stop in between stays short: up to max_stop_minutes (default 30) without charging, or max_charging_stop_minutes (default 120) when a charging session happened during the stop. Overnight stops end a trip.
| Tool | What it returns |
|---|---|
get_trips |
One row per trip: legs, stops and charging stops, distance, driving vs stopped time, moving speed, battery % at start and end, kWh added on the way; filter by date, car, distance, legs, or a place the trip passed through |
get_trip_details |
One trip’s timeline: each drive, and each stop with where and how long, plus kWh and battery % for charging stops |
get_trip_route |
GPS track of a whole trip across its drives, downsampled, with the stop before each leg |
Fast charging (3)
A session is DC when a sample came from a fast charger, or had no AC phases while power flowed; a Tesla-branded fast charger is a Supercharger.
| Tool | What it returns |
|---|---|
get_fast_charging_sessions |
DC sessions with peak power, average power, average power across 20–80% (the fair comparison), minutes from 20% to 80%, whether the battery heater ran, and cost |
get_fast_charging_by_location |
The same per location and charger type: which fast chargers are fastest and cheapest for your car |
get_charging_curve_comparison |
Power against battery % for several sessions, for overlaying their curves |
Charts (8)
show_charging_curve, show_charging_curve_comparison, show_battery_degradation, show_drive_route, show_trip_route, show_activity_report, show_visited_places, and show_efficiency_vs_temperature are the chart versions of their get_* tools. On chart-capable clients they draw an interactive chart in the conversation. On other clients they return the same rows as the get_* tool. The route and trip maps draw the track over a low-detail basemap (Esri gray canvas, light or dark to match the client); the trip map also marks every stop, with charging stops highlighted. Both maps draw an elevation profile under the track; hovering it finds the place on the map. show_activity_report is a monthly or yearly dashboard, show_visited_places maps where the car parks (sized by arrivals, charging places highlighted), and show_efficiency_vs_temperature plots every drive’s consumption against outside temperature with the 5 °C average. Set MAP_TILES=false to keep the maps fully offline.
Custom (2)
get_database_schema— lists all tables, shows the columns of one table, and re-reads the schema when you passrefresh=true.run_sql— runs one customSELECTorWITH … SELECT.
Prompts (11)
Ready-made workflows that name the tools to call, in order: status_report, summarize_driving, analyze_battery_health, analyze_charging, find_anomalies, weather_efficiency, diagnose_sleep, plan_trip, review_trip, monthly_recap, and charging_costs_and_savings. With cost writes on, backfill_costs_from_receipts joins them.
Cost writes (opt-in, off by default)
Set ENABLE_CHARGING_WRITES=true to register set_charging_cost(charging_process_id, cost). It sets the cost of one charging session. This is the same field that the TeslaMate UI edits. A backfill_costs_from_receipts prompt guides the receipt-matching workflow.
Give the database role write access to that single column only:
GRANT UPDATE (cost) ON charging_processes TO teslamate_ro;
This grant is the security boundary. The tool cannot write anything else. On clients with MCP elicitation, the user confirms each write in a dialog. run_sql stays read-only in all cases.
Configuration
The server reads all settings from environment variables. It also reads a .env file. Only DATABASE_URL is required.
| Variable | Default | Purpose |
|---|---|---|
DATABASE_URL |
required | postgresql://user:pass@host:5432/teslamate |
AUTH_TOKEN |
empty | Turns on bearer auth for the HTTP endpoint |
CF_ACCESS_TEAM_DOMAIN |
empty | Cloudflare Access team domain; accepts Access-verified requests |
CF_ACCESS_AUD |
empty | Audience tag of that Access application (set with the above) |
HOST |
0.0.0.0 |
HTTP bind host |
PORT |
8888 |
HTTP bind port |
POOL_MIN_SIZE |
1 |
Minimum pool connections |
POOL_MAX_SIZE |
10 |
Maximum pool connections |
STATEMENT_TIMEOUT_MS |
30000 |
statement_timeout for every query, including reports |
QUERY_TIMEOUT_MS |
5000 |
Tighter statement_timeout for run_sql |
CUSTOM_SQL_ROW_LIMIT |
1000 |
Most rows run_sql returns, whatever its LIMIT |
REPORT_TIMEZONE |
UTC |
IANA timezone for report buckets |
MAP_TILES |
true |
Basemap under the route, trip, and places maps (Esri tiles) |
ENABLE_CHARGING_WRITES |
false |
Registers set_charging_cost |
LOG_LEVEL |
INFO |
Python log level |
DEBUG |
false |
Starlette debug mode; keep off in production |
Generate a bearer token:
uv run teslamate-mcp gen-token
Add your own query
No code change is necessary. The server finds new queries at startup.
-
Put a
SELECTinsrc/teslamate_mcp/queries/your_query.sql. -
Add
your_query.tomlnext to it:name = "get_your_data" description = "What this returns, units, grouping, and available filters." [[params]] # optional: typed tool arguments name = "car_name" type = "string" # string | integer | number | boolean description = "Case-insensitive substring match on the car's name." [[params]] name = "limit" type = "integer" description = "Maximum number of rows returned." default = 10 minimum = 1 maximum = 100 [[output]] # optional: one entry per result column name = "car_name" type = "string" -
Write parameters as
%(car_name)splaceholders in the SQL. Do not build SQL from strings. Cast the first use of each placeholder, for example%(limit)s::int. Write a literal%as%%. The reserved%(tz)splaceholder bindsREPORT_TIMEZONE. -
Restart the server. The registry validates the pair and registers the tool. Run
teslamate-mcp list-toolsto confirm.
The startup validation rejects a query when a declared parameter is missing from the SQL, or when the SQL uses an undeclared parameter.
Security
run_sqlruns in aREAD ONLYtransaction that always rolls back. Timeouts and a row cap apply.- Connect with a
SELECT-only PostgreSQL role, not TeslaMate’s ownteslamateuser. TeslaMate’s Compose makes that user a superuser, and a read-only transaction still lets a superuser read server files throughrun_sql. SECURITY.md has theCREATE ROLEsnippet. - The maps’ basemap tiles are fetched from Esri by the viewer’s browser, which reveals the rough area being viewed.
MAP_TILES=falseturns this off. - The HTTP transport compares bearer tokens with a timing-safe function. Behind Cloudflare Access, it can instead verify the signed
Cf-Access-Jwt-Assertion(signature, issuer, audience, expiry). - Report vulnerabilities through private security advisories. See SECURITY.md.
Development
uv sync # install with dev dependencies
uv run ruff check src tests # lint
uv run ruff format src tests # format
uv run pytest # database tests need Docker or TESLAMATE_TEST_DATABASE_URL (a scratch database)
License
MIT — see LICENSE.
설치
docker run --rm -e DATABASE_URL=... -p 8888:8888 ghcr.io/cobanov/teslamate-mcp:latest설정
{
"mcpServers": {
"teslamate": {
"command": "uv",
"args": ["--directory", "/path/to/teslamate-mcp", "run", "teslamate-mcp", "stdio"]
}
}
}