mtl-c2pa-mcp: Instructions file for Claude Code

mtl-c2pa-mcp

MCP server for reading C2PA content provenance manifests from media files.

Tech Stack

• Language: Python 3.10+
• Package manager: Poetry
• Core library: c2pa-python (official Adobe Rust binding)
• MCP framework: FastMCP (mcp[cli])
• Linter: Ruff
• Test framework: pytest

Commands

Action	Command
Install	poetry install
Run server	poetry run python -m mtl_c2pa_mcp
Test	poetry run pytest
Lint	poetry run ruff check .
Format	poetry run ruff format .

Architecture

• src/mtl_c2pa_mcp/server.py — FastMCP tool definitions
• src/mtl_c2pa_mcp/c2pa.py — Wrapper around c2pa-python Reader: read manifest store, summarize, verify, scan
• src/mtl_c2pa_mcp/__main__.py — Entry point (python -m mtl_c2pa_mcp)

C2PA primer

A C2PA manifest store has the shape:

{
  "active_manifest": "urn:c2pa:<uuid>",
  "manifests": {
    "urn:c2pa:<uuid>": {
      "claim_generator_info": [{"name": "...", "version": "..."}],
      "instance_id": "...",
      "assertions": [
        {"label": "c2pa.actions.v2", "data": {"actions": [...]}},
        {"label": "c2pa.hash.data", "data": {...}},
        ...
      ],
      "ingredients": [...],
      "signature_info": {"issuer": "...", "common_name": "..."}
    }
  },
  "validation_state": "valid",
  "validation_results": {...}
}

Key things to extract for a useful summary:

• Generator: claim_generator_info[0] — who produced the asset (e.g. "Google C2PA Core Generator Library")
• Actions: assertions with label starting c2pa.actions — what was done (created, edited)
• Digital source type: IPTC URI on actions — trainedAlgorithmicMedia means AI-generated
• Watermarks: soft-binding assertions or action descriptions mentioning "watermark" (e.g. SynthID)
• Ingredients: source assets used (parent/component) — empty for purely AI-generated tracks
• Signature: signature_info.issuer — who signed the claim