vehicle-safety-mcp

An MCP server that connects Claude to NHTSA vehicle safety data — VIN decoding, safety recalls, NCAP crash-test ratings, and consumer complaints. Ask Claude "does my car have any recalls?" and get a real answer from the US government's own data.

Built on the free NHTSA public APIs — no API key required.

Why

Vehicle safety data is public but painful: it lives across several government APIs, each response carries 100+ mostly-empty fields, and safety ratings require a two-step lookup nobody remembers. This server does the plumbing so the conversation stays natural:

You: Is a 2020 Honda Civic safe? Anything I should know before buying one?

Claude: (calls get_safety_ratings + get_recalls + get_complaints) The 2020 Civic earned a 5-star overall NCAP rating… it has a handful of recall campaigns worth knowing about… and the component owners complain about most is…

Related MCP server: Vehicle Database MCP Server

Tools

Responses are deliberately trimmed for LLM consumption — raw NHTSA payloads are noisy, so the server keeps the fields that change an answer and drops the rest. Less context burned, better answers.

Quickstart

Requires uv (it manages Python ≥3.11 for you).

git clone https://github.com/basit-khan-abdul/vehicle-safety-mcp.git
cd vehicle-safety-mcp
uv sync

Claude Code

claude mcp add vehicle-safety -- uv run --directory /path/to/vehicle-safety-mcp vehicle-safety-mcp

Claude Desktop

Add to claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "vehicle-safety": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/vehicle-safety-mcp", "vehicle-safety-mcp"]
    }
  }
}

Restart Claude Desktop; the tools appear under the 🔌 icon.

Try asking

• "Decode this VIN: 5UXWX7C5BA"*

• "Does the 2020 Honda Civic have any recalls?"

• "Compare crash-test ratings of a 2021 Toyota RAV4 and a 2021 Honda CR-V"

• "What do owners complain about most on the 2019 Ford F-150?"

Reliability

Every NHTSA call runs with explicit connect and read timeouts and retries only transient failures — 5xx responses and connection/timeout errors — up to 3 attempts with exponential backoff and full jitter; client errors (4xx) fail fast and are never retried. When the upstream is genuinely unreachable, tools return a structured {"error", "detail", "available": false} payload the model can relay honestly, never a raw traceback into the MCP layer. Timeouts, attempt count, and backoff are all environment-tunable (see Tuning). And because the NHTSA APIs are the real contract, a weekly scheduled job runs live smoke tests against them and opens a GitHub issue the moment a field or response shape drifts — so breakage surfaces before users do.

Tests

Two suites, split by what they prove:

• Unit (tests/unit/) — mocked, deterministic, offline. Run on every push/PR across Python 3.11 and 3.12 (CI).

• Live (tests/live/) — real NHTSA API smoke tests. The upstream contract is the product, so a weekly scheduled job (contract-drift) re-runs them and opens an issue if the API drifts — instead of making every push depend on a third-party API.

uv run --extra dev pytest tests/unit     # fast, offline — what CI runs on push
uv run --extra dev pytest tests/live     # hits the real NHTSA APIs
uv run --extra dev pytest -m "not live"  # everything except live

Design notes

• Trimmed responses over raw passthrough. Each NHTSA record is filtered to a curated field list (nhtsa.py). An LLM doesn't need 140 vPIC fields to say "it's a BMW X3".

• Composite tool for the common question. check_vin_recalls chains decode → recall lookup because "does my car have recalls?" is the question real people ask — one tool call instead of two round trips.

• Bounded output everywhere. Ratings capped at 5 variants, complaint narratives truncated at 400 chars, complaint list limited — tool output that scrolls forever helps nobody.

• Resilient by default. Every NHTSA call has explicit connect/read timeouts and retries transient failures (5xx, connection errors, timeouts) up to 3 attempts with jittered backoff — but never retries a 4xx. When the upstream is genuinely down, tools return a structured {"error": …, "available": false} payload the model can relay honestly ("NHTSA data is currently unreachable; try again later") instead of surfacing a raw traceback.

Tuning

The HTTP behaviour is env-configurable (sensible defaults shown):

MCP registries

Built to the Model Context Protocol spec and installable from source (see Quickstart). Browse or publish MCP servers via the community registries:

• Official MCP Registry

• Glama MCP directory

• MCP.so

• PulseMCP

Data source & disclaimer

Data comes from the US National Highway Traffic Safety Administration public APIs and covers vehicles sold in the United States. This project is not affiliated with or endorsed by NHTSA. Always verify safety-critical information with official sources.

License

MIT © Basit Khan