krisp-meeting

Krisp Meeting Skill

Fetch, summarize, and optionally download transcripts from Krisp meetings via the Krisp MCP API.

Authentication

Krisp uses OAuth tokens managed by mcp-remote. Tokens are stored at:

~/.mcp-auth/mcp-remote-{version}/{hash}_tokens.json

Run this script first — it handles finding the token, checking expiry, and refreshing if needed:

python3 ~/.claude/skills/krisp-meeting/scripts/get_token.py

It prints the access token to stdout, or exits with an error message if auth is needed.

If auth fails: Run npx -y mcp-remote https://mcp.krisp.ai/mcp in a terminal — it will open a browser to complete OAuth, then store the token. After that, re-run the skill.

Alternatively, read the token directly:

TOKEN=$(python3 -c "import json; print(json.load(open(__import__('os').path.expanduser('~/.mcp-auth/mcp-remote-0.1.37/e8fe1423b57023b6315cd9029dbc1859_tokens.json')))['access_token'])")

Or find any token file dynamically:

TOKEN=$(python3 -c "
import glob, json, os
files = glob.glob(os.path.expanduser('~/.mcp-auth/mcp-remote-*/*_tokens.json'))
files.sort(key=os.path.getmtime, reverse=True)
print(json.load(open(files[0]))['access_token'])
")

Calling the Krisp MCP API

All API calls follow this pattern:

curl -s -X POST "https://mcp.krisp.ai/mcp" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $TOKEN" \
  -H "Accept: application/json, text/event-stream" \
  -d '{...}' \
  | grep "^data:" | sed 's/^data: //'

Response is SSE — extract the data: line and parse as JSON. The result lives at .result.content[].text (text blocks) or .result.structuredContent (structured data).

Important: Save curl output to a file before parsing with Python to avoid shell encoding issues:

curl ... > /tmp/krisp_result.txt
grep "^data:" /tmp/krisp_result.txt | sed 's/^data: //' > /tmp/krisp_data.json
python3 -c "import json; data = json.load(open('/tmp/krisp_data.json', 'rb').read().decode('utf-8', errors='replace'))"

Available MCP Tools

search_meetings

Find meetings by recency, text, or date range.

{
  "jsonrpc": "2.0", "id": 1,
  "method": "tools/call",
  "params": {
    "name": "search_meetings",
    "arguments": {
      "limit": 1,
      "fields": ["name", "date", "attendees", "speakers", "meeting_notes", "key_points", "action_items"]
    }
  }
}

Key arguments:

• limit / offset — pagination (default limit 10)
• search — text search by topic, name, or attendee
• after / before — ISO 8601 date filters (e.g. "2026-04-06")
• fields — which fields to return

Response includes meeting_id (32-char hex UUID) needed for get_multiple_documents.

get_multiple_documents

Fetch the full meeting document including complete transcript, notes, and action items.

{
  "jsonrpc": "2.0", "id": 2,
  "method": "tools/call",
  "params": {
    "name": "get_multiple_documents",
    "arguments": { "ids": ["<32-char-meeting-id>"] }
  }
}

The response text contains the full markdown document: summary, key points, action items, and the verbatim transcript with speaker labels and timestamps.

Workflow

1. Get token

Read the token using one of the patterns above. If it fails, tell the user to run npx -y mcp-remote https://mcp.krisp.ai/mcp to re-authenticate.

2. Find the meeting

Call search_meetings based on what the user asked:

• Default / "last meeting" → limit: 1 (returns most recent)
• By keyword → search: "keyword"
• By date → after: "YYYY-MM-DD" + before: "YYYY-MM-DD" (same day for a specific date)
• By person → search: "person name"

3. Get full content (if needed)

If meeting_notes.key_points is empty (common for recent meetings still processing), call get_multiple_documents with the meeting_id to get the full document including the pre-computed key points and complete transcript.

4. Format and respond

Standard summary output:

## [Meeting Title]
**Date:** [formatted date + time]
**With:** [speakers other than the user, comma-separated]

### Key Points
- [bullet per key point]

### Action Items
- [ ] [assignee]: [action item]

If action items only: Just show the action items section with the meeting title and date as context.

If --save or "save to desktop": After displaying the summary, write the full transcript to a .txt file:

1. Call get_multiple_documents to get the full transcript text
2. Strip markdown bold (**text** → text) and header markers
3. Filename: [FirstName]_[LastName]_[Topic]_[YYYY-MM-DD].txt saved to ~/Desktop/
4. Confirm the save path and file size to the user

5. Handle edge cases

• No meetings found: Tell the user and suggest a different search term or date
• {{Speaker_2}} placeholders: Krisp uses these when speaker names aren't resolved — replace with the known name if mentioned elsewhere, or use "the other participant"
• Token expired: The script will attempt a refresh automatically. If it fails, prompt the user to re-authenticate via npx -y mcp-remote https://mcp.krisp.ai/mcp