# Constat — 483 Risk Radar (MCP server)

[![Constat MCP — FDA Device Evidence Lifecycle MCP connector – tool definition quality and endpoint health on Glama](https://glama.ai/mcp/connectors/com.healthai/radar/badges/score.svg)](https://glama.ai/mcp/connectors/com.healthai/radar)

<a href="https://glama.ai/mcp/servers/@thehealthai/fda-risk-radar-mcp">
  <img width="380" height="200" src="https://glama.ai/mcp/servers/@thehealthai/fda-risk-radar-mcp/badge" alt="Constat 483 Risk Radar MCP server" />
</a>

[![MCP Queen grade](https://mcpqueen.com/badge/com.healthai/radar.svg)](https://mcpqueen.com/s/com.healthai/radar)

The MCP server behind [**Constat**](https://constat.dev) — FDA & NHTSA
regulatory-risk intelligence for AI agents, over the Model Context Protocol.
Live public regulatory data — recalls, adverse events, warning letters, 510(k)
premarket evidence, postmarket drift signals, reimbursement pathways, and
vehicle safety. **Decision support, not regulatory advice.**

- **Endpoint:** `https://constat.dev/api/mcp` (legacy `https://radar.healthai.com/api/mcp` still serves)
- **Transport:** Streamable HTTP (JSON-RPC 2.0)
- **Auth:** none for anonymous access (10 work units/min and 50/day per
  anonymous source); an optional `X-API-Key` header enables the separately
  assigned metered allowance. Tool calls have documented work-unit costs.
- **Registry:** [`com.healthai/radar`](https://registry.modelcontextprotocol.io) (official MCP registry)

The evidence corpus and tool logic run hosted at the endpoint above. This repo
also ships `server.mjs`, a zero-dependency **stdio bridge** to that endpoint, so
stdio-only MCP clients can use the server like any local one:

```bash
node server.mjs            # stdio MCP server, bridges to constat.dev/api/mcp
```

```jsonc
// e.g. in an MCP client config
{ "mcpServers": { "constat": { "command": "node", "args": ["/path/to/server.mjs"] } } }
```

```bash
docker build -t constat-mcp . && docker run -i constat-mcp   # same, containerized
```

## Tools

| Tool | What it does |
|------|--------------|
| `device_risk_lookup` | FDA compliance risk for a device category by 3-letter product code — recalls, MAUDE trend, warning-letter matches, composite score |
| `firm_compliance_history` | Source-bounded FDA public-record timeline for a device firm — recalls, warning letters, 483s, clearances |
| `watchlist_diff` | Machine-detected FDA public-record changes for monitored product codes since a given date |
| `device_evidence_lookup` | Parsed 510(k) premarket evidence for an AI/ML device, each field with a verbatim source quote + page |
| `evidence_search` | Find AI/ML clearances by product code, panel, applicant, clinical data, sensitivity metric, or PCCP |
| `predicate_chain` | Trace a device's predicate ancestry with each predicate's age at clearance |
| `evidence_cohort_stats` | Reporting-rate stats across the parsed AI/ML corpus — presence figures with denominators |
| `device_postmarket_lookup` | Post-clearance intelligence for one device — recalls, MAUDE trend, letter/483 matches, drift signals |
| `postmarket_search` | Find devices by postmarket criteria — drift signals, recalls in 24mo, rising MAUDE trend |
| `cohort_postmarket_stats` | Postmarket presence rates across the AI/ML cohort, each with its denominator |
| `reimbursement_lookup` | Clearance-to-payment pathway by K/DEN or CPT code — NTAP, Cat I/III, CMS rates, HCPCS, LCDs |
| `reimbursement_search` | Find payment pathways by mechanism, CPT category, NTAP status, applicant |
| `reimbursement_stats` | Mechanism distribution across the reimbursement corpus with dollar ranges |
| `vehicle_risk_lookup` | NHTSA safety history by make/model/year — recall campaigns and complaint stats |

## Quick start

```bash
curl -s https://constat.dev/api/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"device_risk_lookup","arguments":{"product_code":"FRN"}}}'
```

List tools with `{"method":"tools/list"}`. Note the `Accept` header must include
`text/event-stream` (streamable-HTTP requirement).

Each of the 14 tools declares an output schema. Tool calls return structured
content with an explicit `ok`, `not_found`, `invalid_request`, or `unavailable`
status while preserving a text result for older clients.

The tools are observational and non-destructive, but advertise
`readOnlyHint: false` because calls write bounded quota and operational-
telemetry state. They do not publish or modify FDA, CMS, NHTSA, or other
third-party records.

See the [integration guide](https://constat.dev/integrations), [architecture
and trust boundaries](https://constat.dev/architecture), [privacy
notice](https://constat.dev/privacy), and [support page](https://constat.dev/support).

## Related servers

- [Clarity MCP](https://github.com/thehealthai/clarity-mcp) — condition-aware
  ingredient, product & supplement safety (verdict + evidence tier + citation).
- [MCP Queen](https://github.com/mcpqueen/mcpqueen) — the graded MCP registry
  that independently probes and grades this server.

## About

Built by [Health AI](https://healthai.com), the team behind
[Constat](https://constat.dev). Informational only; verify against primary
FDA/NHTSA sources.

## License

MIT — see [LICENSE](LICENSE).