rdfitted/granola

Granola API Skill

Wraps Granola's private REST API to read meeting documents, AI-generated summary panels, and full audio transcripts directly from the server — bypassing the local cache-v6.json which only retains ~3 transcripts at a time.

When to use

• "What did <person> say about X" / "Find quotes from <person>"
• Sentiment / testimonial / praise mining for any client
• Pulling decisions or action items from older meetings
• Cross-meeting analysis (themes across N weeks)
• Anything where /client-meetings returned "transcript not cached" or notes were empty

Do not use for the 3 most-recently-cached transcripts — /client-meetings reads them from cache-v6.json directly without an API roundtrip.

Auth

Uses the WorkOS access token from Granola's local app data (Granola desktop app must be installed and logged in). The Cognito token is rejected. The required client header is X-Client-Version: 6.4.0.

Recent Granola builds encrypt the token store as %APPDATA%/Granola/supabase.json.enc (the plain supabase.json is left behind and goes stale). The script transparently decrypts the .enc file using the per-install DEK from storage.dek, which is wrapped with the Chromium os_crypt master key from Local State (DPAPI-protected on Windows). Plain supabase.json is still used as a fallback if no .enc file exists.

Requires (Windows): pip install -r ~/.claude/skills/granola/scripts/requirements.txt (cryptography + pywin32).

If auth-check fails with 401, the token has expired — open the Granola desktop app to refresh, then retry.

CLI

All operations go through scripts/granola_api.py. Run from any cwd.

# Verify auth
python ~/.claude/skills/granola/scripts/granola_api.py auth-check

# List meetings for a client (slug from CLIENTS index: tog | gvb | ubs | scattered-ashes-maui | totum)
python ~/.claude/skills/granola/scripts/granola_api.py list-meetings ubs --limit 20 --since 2026-01-01

# Fetch the AI summary panel for a single meeting
python ~/.claude/skills/granola/scripts/granola_api.py panels <doc-id>

# Fetch the full transcript for a single meeting (microphone = Ryan, system = others)
python ~/.claude/skills/granola/scripts/granola_api.py transcript <doc-id>

# MINE SENTIMENT across many meetings — pulls transcripts for matching docs and surfaces
# positive turns from non-Ryan voices, with Ryan's preceding turn for context
python ~/.claude/skills/granola/scripts/granola_api.py mine-sentiment ubs --limit 10 --since 2026-01-01 --out output/ubs-sentiment.md

--limit caps how many meetings get transcribed (each transcript fetch is a separate API call — be polite). Default is 10.

Client folder mapping

Hardcoded inside the script. Update there if folders change:

| Slug | Granola folder ID | |---|---| | tog | dc1f2066-591a-4d12-83f1-89308a5086a2 | | gvb | 3fe358e9-0766-424b-8633-5b5d28fdc3ec | | ubs | 80697bb3-3533-48d5-b3f2-a8c4771a072e | | scattered-ashes-maui | edc01423-0020-4e41-a9ee-196c262a65e6 | | totum | c6e71316-151c-42e2-824b-87d154d591a1 |

Output format

mine-sentiment writes a Markdown file (default output/<slug>-sentiment-<date>.md) grouped by meeting, with:

• Date + meeting title + doc ID
• Per-hit: matched keyword, ~150 chars of preceding Ryan context, ~250 chars of the speaker quote
• Coverage stats footer (N meetings scanned, M hits, 0-hit meetings listed)

Wiki integration

After running mine-sentiment, the natural follow-up is appending the surfaced quotes to ~/.ai-docs/wiki/clients/<slug>.md under a "Client Sentiment — Positive Signals" section, then logging the add in ~/.ai-docs/wiki/log.md. See existing TOG/GVB pages for the format.

Reference: known endpoints

POST https://api.granola.ai/v2/get-documents          {limit, offset, doc_ids?}
POST https://api.granola.ai/v1/get-document-transcript {document_id}
POST https://api.granola.ai/v1/get-document-panels    {document_id}

Required headers: Authorization: Bearer <workos.access_token>, X-Client-Version: 6.4.0, Content-Type: application/json, Accept-Encoding: gzip.