OpenDeviceDB

@opendevicedb/mcp

MCP server that lets AI agents (Claude Code, Codex, Cursor, REA…) ask OpenDeviceDB before reverse engineering a device — "has this been done already, is someone on it?" — and record their work after. Stdio transport, official @modelcontextprotocol/sdk.

Requires Node ≥ 22.18 (runs the .ts sources directly, no build step) and a checkout of this repository with npm install done at its root.

Tools

tool	use it to
lookup_device(id?, vendor?, model?, fcc_id?, usb?, ble_name?, zigbee_model?)	check one device: full record, effective status, live claims and a one-line recommendation; says explicitly when nothing is known
search_devices(query, category?, status?, limit?)	explore: ranked matches over names, aliases and identifiers (with the easiest install difficulty when there is one)
claim_device(id? | vendor+model, category?, who, contact, scope, days=30)	announce you are working on it (max 90 days); creates a stub record if needed
publish_findings(id, patch, allow_downgrade?, release_claim_of?)	merge your results into the record (validated, status only moves forward, policy-screened)
request_device(vendor, model, fcc_id?, usb?, ble_name?, zigbee_model?, goal?)	your user wants the device but will not reverse engineer it: the existing request to add a 👍 to, or a pre-filled link to a new public GitHub issue they submit themselves (read-only, never posts)
list_sources()	imported databases with license, snapshot date and record count, plus the curated count

Install routes. A record may carry install: ways to put its open firmware or software on the device (method such as web-serial, ota-exploit, usb-bootloader, pairing; the project's official guide and installer; difficulty; tools; opens_device, mains_powered, brick_risk, reversible/revert, firmware_constraints). lookup_device returns them in record.install and its recommendation tells the agent to walk the user through the easiest one with the official guide (never step by step when a mains-powered device must be opened); search_devices gives the easiest difficulty as install; publish_findings accepts routes in patch.install, merged by method + guide. Nothing is flashed or hosted here.

Writes go to data/devices/<vendor>/<model>.yaml in the local checkout. The server never runs git: each write result names the changed file and gives the next step (git checkout -b … && git add … && git commit … && gh pr create --fill).

Backends (environment)

variable	effect
ODB_REPO=/path/to/opendevicedb	local checkout: reads curated + imported records, writes allowed
ODB_API_URL=https://…	read-only: the site's static API (/api/index.json, /api/devices/<id>.json), cached 10 min; claim/publish return an error asking for ODB_REPO
neither	the checkout this package lives in (same as ODB_REPO pointing at it)

ODB_REPO wins when both are set.

Register it

Replace /path/to/opendevicedb with your checkout.

Claude Code

claude mcp add opendevicedb -e ODB_REPO=/path/to/opendevicedb -- node /path/to/opendevicedb/packages/mcp/src/server.ts

(--scope user to have it in every project.)

Codex (~/.codex/config.toml, or codex mcp add opendevicedb -- node …/server.ts)

[mcp_servers.opendevicedb]
command = "node"
args = ["/path/to/opendevicedb/packages/mcp/src/server.ts"]
env = { ODB_REPO = "/path/to/opendevicedb" }

Cursor (.cursor/mcp.json in a project, or ~/.cursor/mcp.json)

{
  "mcpServers": {
    "opendevicedb": {
      "command": "node",
      "args": ["/path/to/opendevicedb/packages/mcp/src/server.ts"],
      "env": { "ODB_REPO": "/path/to/opendevicedb" }
    }
  }
}

Read-only use without a checkout of your own data: set ODB_API_URL to the site's base URL instead of ODB_REPO.

Agent workflow

1. lookup — lookup_device with whatever is on the label (FCC id, USB vid:pid from lsusb, BLE name) or vendor + model. Follow the recommendation: if open firmware or an SDK exists, use it; if someone holds a live claim, contact them. Nothing known and your user only wants the device to exist? Give them the request link (request / request_url in the lookup result, or request_device) so the interest counts.
2. claim — nothing known (or only part)? claim_device with who/contact/scope before starting, then commit the file and open the PR so the claim is visible to others.
3. work — reverse engineer. Never touch Remote ID, geofencing or other safety functions; never copy firmware images or vendor code (POLICY.md).
4. publish — publish_findings with summary, status, interfaces (link your protocol doc as spec), projects (your code, hosted elsewhere), install routes, verification and provenance.ai_assisted. Pass release_claim_of to end your claim.
5. PR — run the git / gh pr create command from the result. Maintainers review.

SKILL.md is the same workflow as an agent skill, ready to drop into a skills directory (or embed in another tool).

CLI

The same functions from a shell (same env vars), printing the tool result as JSON:

node src/cli.ts lookup --vendor DJI --model "Neo 2"
node src/cli.ts lookup --usb 2ca3:001f
node src/cli.ts search tello --category drone
node src/cli.ts claim --vendor Acme --model "X1" --category toy --who alice --contact @alice --scope "BLE control" --days 30
node src/cli.ts publish acme/x1 --patch '{"status":"partial-protocol","summary":"…"}'   # or --patch-file findings.json
node src/cli.ts sources

Development

npm test -w @opendevicedb/mcp   # node:test; uses temporary fake repositories, plus a stdio end-to-end test
npm run mcp                     # from the repo root: start the server on stdio

Code: AGPL-3.0-or-later (see the repository LICENSE). SKILL.md alone is CC0-1.0 so other tools can embed it.