Record the meeting. Then hand it to the AI you already use.
Parley does four things: record, transcribe, analyze, search. It captures both sides of a call, gives you a speaker-labelled transcript live, writes a debrief when the meeting ends, and makes every word you have ever recorded searchable.
Then it opens all of it over a local MCP server — 48 tools — so Claude, Claude Code, Cursor or any other MCP client can read, search, file and re-analyze your whole meeting history without a copy-paste in sight.
- 🎙️ Both sides, live — your mic and the meeting's system audio, transcribed and diarized as you talk.
- 📼 A debrief that stays — commitments, action items, findings on a timeline, and a delivery scorecard. Generated once, saved with the recording.
- 🔎 Searchable forever — full-text across every transcript line and every conclusion the analysis drew.
- 🔌 Open by protocol — anything the app can do to your library, an MCP client can do too, including owning the analysis outright.
Local-first, bring your own keys. Audio and transcripts go directly to the STT and LLM providers you configure (Claude, OpenAI, Gemini, Soniox, Deepgram, …). No Pathors proxy, no telemetry, everything stored on your machine.
Note
macOS only (for now). Parley captures the other side of a call through a Core Audio process tap, which has no equivalent on other platforms. There are companion apps for iPhone and Android that record in-person meetings and sync to the same account.
Download the latest build from the Releases page, open the .dmg, and drag Parley into Applications. Builds are signed and notarized — no Gatekeeper hoops.
Then paste your API keys in Settings: one STT provider for transcription, one LLM for analysis and Ask. Or sign in and use the hosted providers with no key to manage.
Starting a meeting is one button. Parley captures both sides — your mic and the meeting's system audio — and transcribes them live, diarized as me / them, with editable speaker names.
Describe the meeting in its context field — who is in the room, what is at stake, what you want the analysis to watch for — and that description goes verbatim into every analysis prompt for the recording. It is also the field an MCP client writes to when it wants to re-aim the analysis.
Beside the transcript sits the coach feed: evaluation alerts from your own templates (negotiation risk, qualification gaps, red flags, or whatever rubric you wrote), each with a drill-down into how to reply, plus an agenda checklist that ticks itself off as the conversation covers it. Ask anything from the input bar and get answers grounded in the conversation so far.
Stopping a meeting lands on its debrief. Any recording — just finished, pulled from history, dragged in as an audio file, or imported as a .txt transcript — opens in two views:
- Report — one scroll: the debrief with clickable timestamps, both sides' commitments, action items, and your delivery scorecard (measured pace, talk share, filler sounds).
- Replay — the full player: scrub to any moment and re-run the analysis as of that point, click through findings on the timeline, and ask anything about the call from a drawer that follows you across tabs.
Everything generates once and saves with the recording — reopen it a month later and the whole report loads instantly, no extra LLM calls. LLM speaker re-attribution fixes diarization drift by conversational context.
One customer, one folder. Every recording is filed in a folder — moved from the library, from the replay titlebar, or by an MCP client — so the calls with one customer sit together instead of scattering across a flat list. Sign in and a folder can live on the shared organization side instead of your personal one.
search_meetings runs full-text across everything you have recorded: every transcript line (each hit carrying a seek target) plus every brief, finding, action item and meeting context. Scope it to one folder to ask a question about one customer.
Parley runs a local Model Context Protocol server. Point your MCP client at it — the app has to be open — and it gets the live meeting, every saved recording, and the whole filing system:
claude mcp add --transport http parley http://127.0.0.1:3011/mcpThe exact endpoint, a ready-to-paste command, and a live view of what your client is doing are in Settings → MCP Server. 48 tools, in five groups:
| Group | What it reaches | Tools |
|---|---|---|
| The meeting happening now | What is on screen, what has been said, the checklist beside it | get_app_context, get_focused_content, get_transcript, list_todos, add_todo, check_todo, list_evaluations |
| Everything you have recorded | Full-text search and full reads across history | search_meetings, list_recordings, get_recording, rename_recording |
| The library, organized | Personal folders, filing, bulk text import, deletion | list_folders, create_folder, rename_folder, delete_folder, move_recording_to_folder, import_transcript, delete_recording |
| Your team's shared space | The org side, with the same permissions the app enforces | list_orgs, list_org_recordings, list_org_folders, create_org_folder, share_recording_to_org, move_org_recording_to_folder, copy_org_recording_to_personal, delete_org_recording |
| Hand over the analysis | Writing findings, action items and the debrief back into the app | set_recording_analysis, update_recording_meta, list_findings, set_findings, update_finding, upsert_eval_template, upsert_todo_template |
Letting an external AI own the analysis. Turn off Auto-analyze recordings in Settings and a finished or imported recording stays unanalyzed on purpose. An agent polls list_recordings with since / the analyzed flag, reads the transcript, and writes its own findings, action items and brief back with set_recording_analysis — stamping author so its findings stay distinguishable from Parley's own pass. Manual regeneration still works either way.
Conversation content is sensitive, so Parley runs straight from your machine:
- Direct connections — audio and transcripts go only to the providers you configure, under your own keys.
- Local storage — recordings, transcripts, and templates stay in your local app directory.
- No telemetry — nothing tracked, collected, or uploaded.
- Voice typing — system-wide push-to-talk dictation in any app, using your configured STT provider. Hold a key (default Option+Space), speak, release — the text pastes into the frontmost app.
- Import what you already have — drag in an audio file, or import a pile of
.txttranscripts (speaker labels and[HH:MM:SS]timestamps are auto-detected) straight into the right folder. - Reusable playbooks — evaluation and checklist templates: MEDDICC, negotiation terms, interview rubrics, diligence questions, or your own.
- Traditional Chinese — full zh-TW UI and on-the-fly conversion of transcribed text.
Requires Rust (stable) and Bun (or Node.js):
git clone https://github.com/pathorsAI/parley.git
cd parley
bun install
bun run tauri devBefore opening a PR, bunx tsc --noEmit and bunx vitest run must both pass. See CLAUDE.md for the repository conventions.
Contributions are welcome! See CONTRIBUTING.md for how to report bugs, suggest features, and submit pull requests.
Licensed under the Apache License 2.0. Copyright 2026 Pathors AI.