turnwise

An MCP server that scores dialogue transcripts for bridging quality.

Pre/post surveys ask participants how they felt. turnwise reads what they said, and gives each speaker six scores.

### Concession: 4.00, confidence 0.75

Three markers across 175 words, granting a point to the other side twice and
stating a changed view once, for a weighted density of 2.40 per 100 words.

> [turn 11] **Sam:** "Although you got me on the transportation contract."
> [turn 11] **Sam:** "But I'm less comfortable than I was ten minutes ago."

extract_evidence returns the pattern matches with the rule that fired and the turns where nothing matched, leaving you to decide whether a match means what the rule assumes.

Install

npm install && npm run build

Add this to claude_desktop_config.json, then restart Claude Desktop:

{
  "mcpServers": {
    "turnwise": {
      "command": "node",
      "args": ["/absolute/path/to/turnwise/dist/index.js"]
    }
  }
}

Related MCP server: LongBook Verifier

Tools

The parser detects WebVTT, SRT, Otter, Rev, plain Speaker: text, and JSON.

Indicators

Follow-up questions and reciprocated disclosure are computed from turn structure, and explain_indicator covers the rest.

Audio

pip install mlx-whisper        # Apple Silicon, fastest
pip install -U openai-whisper  # cross-platform
brew install whisper-cpp       # no Python
brew install ffmpeg            # video and multitrack

transcribe_audio returns text without speaker labels, because this tool does no diarization.

Changing one recording setting fixes this: in Zoom, turn on Settings > Recording > Record a separate audio file for each participant, then point transcribe_session at the folder. Each file holds one person, so attribution is exact and the filenames become the speaker names.

Limits

Read docs/METHODOLOGY.md before putting these numbers in a report.

Validation against trained human coders is still outstanding, so treat these as evidence-linked observations. Contempt comes in low because tone carries most of it and a transcript has none, and sarcastic agreement will score as ordinary agreement. Nobody assigns participants to workshops at random, so a difference between groups describes those groups.

Development

npm test          # 32 unit tests
npm run calibrate # 22 minimal pairs the instrument must order correctly
npm run smoke     # exercises every tool over MCP

Run calibrate after any lexicon edit, since it also tracks the adversarial cases that pattern matching cannot solve.

Editing src/lexicon.ts changes the instrument, so bump RUBRIC_VERSION when you do, or year-over-year comparisons quietly stop meaning anything.

Apache-2.0.