Tagging pipeline

Postal's tagging model proposes genres, moods, instruments, vocal types and more for every track. This recipe runs it across a library and puts a human in the loop only where it matters.

Steps

1. Find untagged tracks. Page GET /tracks with fields=id,name,audio_profile and pick tracks whose profile has no genres.
2. Queue analysis. Send ids in batches to POST /tracks/auto-tag/batch.
3. Wait. Handle track.audio_profile_completed webhooks, or poll GET /tracks/suggested-tags until analysisStatus is completed.
4. Apply or review. For tracks where intelligence.ready is true and the suggestion looks complete, call POST /tracks/apply-suggested-tags. Route the rest to a review queue.
5. Write corrections. When a reviewer changes something, send the full corrected category with PATCH /tracks/tags. Corrections are recorded as training feedback.

Scopes: read:tracks, update:tracks.

Validating tag names

Load GET /taxonomy once at startup. Reject any tag a reviewer types that is not in it, or map it to the nearest valid name before writing.

Throughput

Analysis runs in Postal's queue and can take from seconds to a few minutes per track depending on load. Queue in batches of a few hundred and let webhooks drive the next step rather than polling every track.

In an assistant

The MCP wait_for_auto_tag tool collapses trigger, poll and read into one call, and apply_suggested_tags commits. See the tool reference.