Open-source audio provenance infrastructure with an MCP adapter so AI assistants can run auditable provenance analysis on real-world AI audio (local workspace files): inspect → verify credentials → distribution-stress → report.
python3 -m venv .venv && source .venv/bin/activate
pip install -e packages/audio_prov -e packages/audio_prov_mcp
python scripts/generate_fixtures.py
audio-prov check
audio-prov run provenance-analysis@1 --asset tone-wav
audio-prov batch --json # sync: all files in workspace/
audio-prov batch --async --json # async: returns batch_id, poll status
audio-prov sign-c2pa your-track.wav # embed C2PA (dev cert by default)See docs/mcp-setup.md for Claude Desktop configuration.
python scripts/wire_claude_desktop.py # macOS: merge MCP config, then restart Claude
cp ~/Downloads/your-ai-track.mp3 workspace/In Claude: use MCP list_workspace_files → register_workspace_file → tool analyze_ai_audio (not provenance_run with prompt name).
For large catalogs, use analyze_workspace_async then poll get_batch_run(batch_id).
- Workspace-first: analyze your own MP3/M4A/WAV exports in
workspace/ - Multi-signal reports: structural, verified, simulated, inferred (tag hints + watermark stub)
- MCP tools:
register_workspace_file,analyze_ai_audio,analyze_workspace_async,sign_c2pa_manifest, etc. - Distribution presets:
aac128,aac64,mp3_128,loudnorm_-14,copy - Pluggable verify: demo sidecar manifests + optional C2PA via
c2patool - C2PA signing: dev cert by default; operator keys via
AUDIO_PROV_C2PA_PRIVATE_KEY/AUDIO_PROV_C2PA_SIGN_CERT
Analysis reports technical evidence only. Absent credentials do not prove synthetic origin. Demo sidecar manifests and default C2PA signing use development keys only.
MIT