voice-operations
Skill4 starsChanged 4 months agoMITView on GitHub
Drive the `openwhispr` MCP — read OpenWhispr voice transcripts (text + `.webm` audio) and own the idempotency watermark. `list_transcripts` returns only UNPROCESSED notes; `mark_processed` (LAST, after a note is fully handled) advances the watermark so it is never re-emitted. Use for any voice-note read, audio fetch, or processing loop.
In this repository17 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- mcp/CLAUDE.mdCLAUDE.md
- add-an-addonSkill
- backup-recoverySkill
- body-mcp-setupSkill
- cos-setupSkill
- cos-upgradeSkill
- debug-cowork-mcp-issuesSkill
- fitness-mcp-setupSkill
- guard-setupSkill
- hub-handoverSkill
- mcp-bridge-setupSkill
- nutrition-mcp-setupSkill
- openwhispr-mcp-setupSkill
- setup-vaultSkill
- spoke-setupSkill
- whatsapp-mcp-setupSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 74
If list_transcripts is stuck on fixtures instead of your real notes, that's OPENWHISPR_FIXTURES set — see /openwhispr-mcp-setup.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Voice operations — read OpenWhispr notes & own the watermark
The openwhispr MCP exposes your OpenWhispr voice transcripts (a local SQLite DB + per-note
.webm audio) read-only, and owns the watermark that makes voice processing idempotent. It is
the read + watermark layer; it does not route anything (turning a note into a vault page or a board
case is /second-brain-ingest's
job). Four tools: list_transcripts, get_transcript, get_watermark, mark_processed.
The tools
list_transcripts(limit?, since?, includeProcessed?)— lists{ id, created, preview }, with a🎙 <file>marker per note's audio. By default returns only UNPROCESSED notes (those sorting after the watermark).includeProcessed: trueignores the watermark and returns everything (browse mode).sincefilters to notes created strictly after a time (interpreted on the local clock — a trailingZ/offset is ignored).limitcaps the count: when the watermark is honored it keeps the oldest unprocessed notes (so a backlog drains in order); when browsing it keeps the most-recent. Watch for the integrity flags⚠ audio file missingand⚠ orphan audio (no transcript row).get_transcript(id)— the fulltextplus metadata:created,has_audio,audio_path(the absolute path to the.webm, so you can fetch the real audio),audio_duration_ms,provider,model,status, andaudio_missing: trueif the recording is gone. Also reports whether the note is already processed. An unknownidreturns a clean tool error.get_watermark()—{ id, created }of the last note marked processed (bothnullif the loop has never run), plus the state-file path. Use it to see where the cursor sits.mark_processed(id)— advances the watermark to that note's{ id, created }. After this,list_transcriptsskips that note and anything older. This is the only state-changing tool — the whole reason the server exists (OpenWhispr has no native "mark read").
The watermark is the point (idempotency)
mark_processed is the only write. The watermark is a single { id, created } JSON file the server
owns (state/watermark.json, override OPENWHISPR_STATE); list_transcripts returns only notes sorting
strictly after it (by created, with id as the tiebreaker for monotonic ids like
vn-2026-05-29-1705-…).
So the correct way to process notes is the loop:
list_transcripts→ the unprocessed notes (the watermark already excludes anything handled).- For each,
get_transcript(id)→ the full text (andaudio_pathif you need the audio), then fully handle it (e.g. hand it to/second-brain-ingestto route knowledge → vault and action → board). mark_processed(id)— LAST, only after the note is fully handled.
Advance the watermark LAST. Because mark_processed runs only after a note is done, the loop is
at-least-once: crash mid-loop and the next run simply reprocesses from the last successfully-marked
note forward — nothing is skipped. Marking before you finish would silently drop a note. With no
watermark file yet, nothing is processed, so every note is "new".
Reading the audio, not just the text
get_transcript gives audio_path — the absolute path to the note's .webm. The DB is the source of
truth for text; audio lives as audio/OpenWhispr-<timestamp>-<id>.webm (the integer before .webm is
the transcript id). If audio_missing is set (or list_transcripts flagged ⚠ audio file missing),
the DB claims audio the file no longer has — handle the text alone.
Where the notes come from (source resolution)
The server resolves its source in order — real store first, degrades gracefully:
OPENWHISPR_FIXTURES=<dir>—*.json/*.mdfixture files (an explicit test override; highest precedence).- SQLite (the real store) —
OPENWHISPR_DB, or the auto-detected macOS path~/Library/Application Support/open-whispr/transcriptions.db, read read-only via thesqlite3CLI (WAL-safe — fine while the app is running). The production default. openwhispron PATH — legacy CLI fallback.- None — every tool returns a clean error telling you to set
OPENWHISPR_DB/OPENWHISPR_FIXTURES.
Soft-deleted rows (deleted_at) are hidden; every other note is kept (even non-completed / empty-text
ones) so a recording is never silently dropped — status / has_audio are surfaced in metadata instead.
If list_transcripts is stuck on fixtures instead of your real notes, that's OPENWHISPR_FIXTURES set —
see /openwhispr-mcp-setup.
Practical notes
- Read-only on your notes. This MCP never writes the transcripts DB or the audio; the only thing it writes is its own watermark file. It cannot delete or edit a note.
- Routing is not this skill's job. This is the read + watermark layer; classifying a note and writing
the vault/board is
/second-brain-ingest(the router) + theboardMCP. Keep the two concerns separate — drive the reads/watermark here, hand the content to the router. - Setup vs operation. Standing the server up on a new machine (the launchd bridge on
:8002, the Cowork stdio entry,OPENWHISPR_DB) is/openwhispr-mcp-setup; this skill is for using it once it's wired.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

