Turnwise is an MCP server that scores dialogue transcripts for bridging quality with per-speaker indicators and evidence.
Score conversations: Get 0–4 scores for receptiveness, perspective-taking, contempt, curiosity, concession, and personal disclosure, with quoted evidence, confidence, and airtime metrics.
Extract evidence: Review raw pattern matches and unmatched turns to adjudicate sarcasm, quotes, or missed behavior.
Parse transcripts: Normalize WebVTT, SRT, Otter, Rev, plain Speaker: text, and JSON formats, and verify speaker detection.
Compare conversations: Put two transcripts side by side with per-indicator differences for before/after or group comparisons.
Analyze cohorts: Aggregate many scored conversations into means, spread, and effect sizes for funder-ready reports.
Explain indicators: Get the full rubric, thresholds, limitations, and research citations for any indicator.
Transcribe audio: Convert a single recording to text locally with no speaker labels.
Transcribe sessions: Merge per-participant recordings into one speaker-labeled transcript for exact attribution.
Check audio support: See which local speech-to-text backends are installed before transcribing.
Enables processing Zoom meeting recordings (multitrack per-participant audio files) into a speaker-labeled transcript for dialogue analysis.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@turnwise Score this conversation transcript for bridging quality"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
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.
npm install && npm run buildAdd 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: tts-audio-mcp
Tool | Purpose |
| Per-speaker scores with evidence. |
| Matches for you to adjudicate, plus unmatched turns. |
| Check speaker detection before scoring. |
| Two sessions side by side. |
| Aggregate many conversations, with effect sizes. |
| Full rubric, limits, and citations for one indicator. |
| Local speech-to-text for one recording. |
| Multitrack folder to one speaker-labeled transcript. |
| Which local backends are installed. |
The parser detects WebVTT, SRT, Otter, Rev, plain Speaker: text, and JSON.
Indicator | Direction | Grounded in |
Receptiveness | higher is better | Yeomans, Minson, Collins, Chen & Gino (2020) |
Perspective-taking | higher is better | Kalla & Broockman (2020, 2023) |
Contempt | lower is better | Gottman & Levenson (1992, 2002) |
Curiosity | higher is better | Huang, Yeomans, Brooks, Minson & Gino (2017) |
Concession | higher is better | Hussein & Tormala (2021); Deutsch (1973); Fisher & Ury |
Personal disclosure | higher is better | Kubin, Puryear, Schein & Gray (2021); Broockman & Kalla (2016) |
Follow-up questions and reciprocated disclosure are computed from turn structure, and explain_indicator covers the rest.
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 multitracktranscribe_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.
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. Participants choose which workshop to attend, so a difference between two groups may come from who signed up rather than from anything the program did.
npm test # 32 unit tests
npm run calibrate # 22 minimal pairs the instrument must order correctly
npm run smoke # exercises every tool over MCPRun 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 stop meaning anything.
Apache-2.0.
Combine scores from multiple conversations into one report: per-indicator means, spread, and optional group comparison with effect size. This is the tool that produces something a funder can read. Pass the structured score objects returned by score_conversation.
Inputs: scores (required), group_by, indicators.
Report which local speech-to-text backends and media tools are installed, with install instructions if none are. Use this before attempting transcription so a missing dependency surfaces as a clear answer.
Inputs: no named parameters.
Score two transcripts and lay them side by side with the difference on each indicator. Use for before/after comparisons: the same group early and late in a program, one workshop format against another, or a facilitator's first session against their tenth.
Inputs: label_a, label_b, indicators, transcript_a, +3 more.
Return the full rubric for one indicator: what it measures, what counts as evidence, what does not count, the 0-4 anchors, the thresholds, what the score cannot tell you, and the research behind it. Use whenever someone asks what a score means or challenges one.
Inputs: indicator (required).
Return every pattern match the instrument found, with the quote and the rule that fired, plus the turns where nothing matched. Use this when the reading matters more than the number: you confirm or reject each candidate yourself, catching sarcasm and quoted speech that a word list scores wrong, and catching behavior in the unmatched turns that no pattern covers. This is the tool for producing a defensible qualitative summary.
Inputs: meta, format, indicators, transcript, +2 more.
Normalize a transcript into speaker turns and compute structural metrics: turn counts, airtime share per speaker, airtime concentration, and question rate. Run this when a transcript's format is uncertain, to confirm speakers were detected correctly before scoring.
Inputs: meta, format, transcript, speaker_aliases, +1 more.
Score a conversation for bridging quality. Returns a 0-4 score per speaker per indicator (receptiveness, perspective-taking, contempt, curiosity, concession, personal disclosure), each with quoted evidence anchored to turn numbers, a confidence value, and deterministic airtime metrics. Set arc_segments to 3 to see how the conversation changed from opening to close.
Inputs: meta, format, indicators, transcript, +3 more.
Transcribe a single audio or video recording to WebVTT using a local speech model. Audio stays on this machine. A mixed track cannot be reliably split by speaker, so the result omits speaker labels and cannot produce per-speaker scores. For those, record one file per participant and use transcribe_session.
Inputs: file_path (required), model, backend, language.
Transcribe a folder of per-participant recordings and merge them into one speaker-labeled transcript ordered by timestamp. Each file contains exactly one person, so speaker attribution is exact. This is the recommended path for anything that needs per-speaker scores. In Zoom, enable Settings > Recording > 'Record a separate audio file for each participant'. Speaker names come from the filenames.
Inputs: directory (required), model, backend, language, +1 more.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 9 tools
Each tool targets a distinct stage in the analysis pipeline, from environment checks to transcription, parsing, scoring, evidence extraction, comparison, cohort analysis, and rubric explanation. No two tools share the same purpose, and descriptions clearly differentiate similar operations like transcribe_audio vs transcribe_session.
All tool names follow a consistent verb_noun snake_case pattern (e.g., check_audio_support, transcribe_session, score_conversation). This makes the action and target immediately clear, and there are no stylistic deviations.
With 9 tools, the set is well-scoped and each tool earns its place. The count fits within the ideal 3-15 range and covers the full workflow without redundancy.
The server covers the entire lifecycle from pre-transcription checks to group-level analysis and rubric explanation. There are no obvious gaps: transcription, parsing, scoring, qualitative evidence review, comparison, and cohort aggregation are all represented.
MCP server providing access to the Scorecard API to evaluate and optimize LLM systems.
MCP server for RiverScript, an AI transcription platform - fetches transcripts shared via a link.
An MCP server that provides tools to discover and retrieve podcast episodes transcripts.
An MCP server that provides congressional transcripts
