c2pa-mcp

An MCP server that verifies C2PA Content Credentials and returns an LLM-ready verdict.

Point any MCP client (Claude Desktop, Claude Code, Cursor, ...) at a local file or a URL and get back a plain-language answer: is this image/video/audio trusted, valid, tampered, or unsigned? Who signed it? Is it AI-generated? What's its edit history and provenance lineage?

Built by c2paviewer.com. Verification runs locally on the official C2PA Rust engine via @contentauth/c2pa-node. Files never leave your machine.

Read-only. This server verifies and inspects Content Credentials. It does not sign or create them.

Install

No global install needed. Add it to your MCP client config and it runs via npx:

{
  "mcpServers": {
    "c2pa": {
      "command": "npx",
      "args": ["-y", "@c2paviewer/c2pa-mcp"]
    }
  }
}

• Claude Desktop: Settings → Developer → Edit Config, add the block above.

• Claude Code: claude mcp add c2pa -- npx -y @c2paviewer/c2pa-mcp

• Cursor / others: add the same mcpServers entry to the client's MCP config.

Requires Node.js 18+.

Related MCP server: MCP Content Credentials Server

Tools

Each verify tool returns a human-readable summary plus a structured digest:

{
  "verdict": "invalid",            // trusted | valid_untrusted | valid_trust_unknown | invalid | no_credentials
  "summary": "Content Credentials are INVALID: an integrity or signature check failed ...",
  "signer": { "name": "Example Signer", "trusted": false },
  "aiGenerated": { "isAI": true, "tools": ["DALL-E"], "digitalSourceTypes": ["...trainedAlgorithmicMedia"] },
  "provenance": [ { "depth": 0, "title": "This file", "relationship": "This file", "verdict": "invalid" } ],
  "edits": [ { "label": "Created", "agent": "Photoshop", "when": "...", "detail": "" } ],
  "watermarks": [ { "kind": "synthid", "assertionLabel": "...", "algorithm": "" } ],
  "issues": [ { "code": "assertion.dataHash.mismatch", "severity": "error",
               "explanation": "The media content was changed after it was signed. ..." } ],
  "trust": { "evaluated": true, "listSource": "https://.../C2PA-TRUST-LIST.pem" }
}

Pass "includeRaw": true to also get the full raw manifest store.

Trust list

To report a signer as trusted (not just cryptographically valid), the server checks the signing certificate against two trust lists, fetched live and cached (24h TTL) so decisions stay current without a release:

• the official C2PA Conformance Program list (going-forward signers), and

• the legacy CAI Interim Trust List (frozen Jan 2026, but still the only list that recognizes pre-conformance signers — Adobe, Leica, Truepic, Canon, Samsung, and most real-world content today).

Both are on by default; without the ITL, mainstream signed content would read as valid_untrusted.

If the trust lists can't be fetched, the server degrades loudly: verification still runs, but the verdict becomes valid_trust_unknown and trust.evaluated is false with a reason. It never silently treats an unknown signer as trusted, and never silently uses a stale snapshot.

Environment overrides:

Security

• Local processing. Files are read and verified on your machine; nothing is uploaded.

• SSRF-guarded URL fetching. verify_c2pa_url allows only public https, pins the resolved IP against DNS rebinding, re-validates every redirect, sends no cookies or auth, and enforces a content-type allowlist plus size cap.

• Local path safety. The file/scan tools reject remote file:// and UNC paths, resolve symlinks before the allowed-roots check, and return a generic error for inaccessible paths.

• Set C2PA_ALLOWED_ROOTS in any untrusted context. The file/scan tools expose the host filesystem by design, so a prompt-injected model could point them anywhere. Confining them to a specific directory is the single most important hardening setting.

• Trust integrity. Trust lists are fetched over HTTPS (integrity rests on TLS and the upstream sources); the cache is per-user (0700/0600, ownership-checked) and bound to the configured URL set.

Limitations

• Experimental. Not legal evidence. C2PA tooling and trust infrastructure are still evolving. Do not rely on these verdicts for legal, compliance, or safety-critical decisions.

• Watermarks are reported as declared, not pixel-verified. A synthid entry means the manifest declares a SynthID watermark; confirming the signal in the pixels requires the vendor's detector.

• AI-generation reflects what the manifest declares via IPTC digitalSourceType. Absence of an AI declaration is not proof the content is not AI-generated.

Development

npm install
npm run build
npm test          # builds, then runs unit + end-to-end tests (network needed for the trust list)

License

Source code is dual-licensed under MIT or Apache-2.0, at your option.

The test images under test/fixtures/ are redistributed unmodified from c2pa-org/public-testfiles and are licensed separately under CC BY-SA 4.0.

Note: if you publish under an unscoped name instead of @c2paviewer/c2pa-mcp, change name in package.json and the args in the install block above; nothing else depends on the package name.