meetey
Allows recording audio from Zoom meetings and generating structured notes and transcripts via local transcription.
Click on "Install 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., "@meeteystart recording my Zoom meeting"
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.

Meetey
Join a meeting, run /meetey start, and get back dated notes where every decision and action item carries the [mm:ss] it came from. Audio comes off the meeting app through macOS ScreenCaptureKit, whisper.cpp does the transcription, and Claude writes the summary. Screen content — slides, screen shares, code — is opt-in per meeting, with text recognized on-device. If you'd rather not remember to hit record, Meetey can watch for meetings and ask.
Recording and transcription happen entirely on your machine. No subscriptions, no third-party meeting bots, no cloud transcription service. See What stays local for exactly what is and isn't sent anywhere.
Install
Requires macOS 13+, Node.js 18+, and Xcode Command Line Tools.
npx jej2k5/meetey installThis will:
Install whisper-cpp via Homebrew (if not already present)
Download the Whisper base model (~142 MB) to
~/.meetey/models/Build and sign the Swift capture binary
Register the MCP server in Claude Code
Install the
/meeteyskillAdd shortcuts inside Claude Code (
Ctrl+Shift+R/Ctrl+Shift+S)
After installation, open System Settings → Privacy & Security → Screen Recording and enable your terminal app. Then start a new Claude Code session.
Related MCP server: Voice MCP
Usage
In Claude Code
/meetey start Detect running meeting apps and start recording
/meetey stop Stop recording, transcribe, and save notes
/meetey status Check whether a recording is active
/meetey watch Ask to record meetings automatically (off by default)
/meetey list Browse past meetings
/meetey show Open one meeting's notes
/meetey search Find a moment across every meeting
/meetey delete Remove a recording
/meetey doctor Check the install/meetey start asks whether to also capture screen content. Answer up front to skip the question:
/meetey start --video Audio + screen keyframes
/meetey start --audio-only Audio onlyOr use the shortcuts:
Shortcut | Action |
|
|
|
|
These are shortcuts inside Claude Code, not system-wide hotkeys — they only fire while Claude Code is the focused window. To stop a recording from anywhere, use the menu bar item.
While a recording is running
A menu bar item appears for as long as Meetey is recording, showing what it's recording and for how long. Stop Recording ends it without switching back to Claude Code.
◉ 12:04 ⌄
┌──────────────────────────────┐
│ Weekly Sync — Google Meet │
│ Audio only · 12:04 │
│ ────────────────────────── │
│ Stop Recording │
└──────────────────────────────┘The dot is red while recording, and it tells you what is being captured: Audio only and Audio + screen are different things to have agreed to, and they never look the same.
If Meetey thinks your call has ended, it says so before acting — and lets you overrule it:
◌ ending soon ⌄
┌──────────────────────────────┐
│ Weekly Sync — Google Meet │
│ Call may have ended · 8:12 │
│ ────────────────────────── │
│ Keep Recording │
│ Stop Recording │
└──────────────────────────────┘It's there partly as a control and partly so you can always tell at a glance that recording is happening — every other signal Meetey gives you is a single moment, and this one lasts the whole meeting.
Stopping saves the audio; it doesn't write the notes. A notification says so, and /meetey stop in Claude Code picks the recording up and writes it up.
Workflow
Join your meeting in Chrome (Google Meet / Teams web), Zoom, or Microsoft Teams
Run
/meetey start— Claude detects the running app, asks whether to capture screen content, and begins recordingWhen the meeting ends, run
/meetey stop— Claude transcribes the audio, reads any screen content, and produces:
## Q3 Roadmap and API Cutover
**Recorded:** Fri 1 Aug 2026, 2:02–2:47 PM · 45 min
### Summary
...
### Key Decisions
- **[12:04]** Ship v1.1 without the motion heuristic; revisit next cycle
### Action Items
- [ ] **[31:17]** Draft the migration doc — Priya · due Friday
### Screen Content ← only when screen capture was used
- **[04:12]** Burndown chart, actual tracking ~11 days behind plan
### Transcript
Full transcript: `2026-08-01-1402-q3-roadmap-transcript.md` · 6,180 wordsEvery claim carries the timestamp it came from. If you doubt a decision, jump straight to that moment in the transcript instead of searching thousands of words.
Notes are written to ~/.meetey/recordings/ as YYYY-MM-DD-HHMM-<slug>.md — chronologically sortable, so ls is a meeting history. The transcript lives in a sibling -transcript.md file rather than inline, so the notes stay scannable.
If you forget to stop
Forget step 3 and the recording still ends itself — when your call ends, or when the app quits or the captured window closes. Run /meetey stop afterwards anyway to write the notes; it picks the finished recording up and tells you why it ended.
Ending on the call itself is worth explaining, because it works differently than you'd expect. Meetey watches whether the meeting app is holding the microphone, which apps release when you hang up (muting yourself doesn't count — that's just a switch). But that clue is occasionally wrong, and stopping a live meeting by mistake loses audio you can't get back. So instead of stopping, it notes the time and keeps recording. If the call turns out to still be going, the note is thrown away and nothing is lost. Only after ten quiet minutes — and only if nobody spoke during them — does it stop, and then it cuts the file back to the moment you actually hung up. The waiting never reaches your transcript.
When the transcription is unreliable, the notes say so at the top instead of summarizing noise with a confident voice:
> ⚠️ Transcription quality: poor — 80% of segments were silence markers,
> fragments, or repeats. Summary may be incomplete; ggml-base.en.bin was used.That assessment is computed from the transcript itself (fragment rate, whisper looping, speech pace against spoken time), not guessed.
Supported apps
App | Audio captured | Screen captured (opt-in) |
Chrome | All Chrome audio | One chosen window, or everything Chrome displays |
Zoom | Zoom meeting audio | The Zoom window, including screen shares |
Microsoft Teams (desktop) | Teams meeting audio | The Teams window, including screen shares |
Chrome note: audio is captured at the process level — all Chrome audio, not just the meeting tab — so mute any other tabs playing audio before starting. For screen capture you can narrow to a single window, which excludes the app's other windows and its notification banners. You cannot narrow to a tab: macOS offers no tab-level capture, so whatever tab that window is showing is what gets captured.
Screen capture
Meetey can also capture what was on screen — slides, screen shares, code, diagrams — and fold it into the notes.
It is off by default and opt-in per meeting. Claude asks every time you run /meetey start; there is no setting that turns it on permanently. Skip the question with /meetey start --video or /meetey start --audio-only.
/meetey start --video Record audio + screen keyframes
/meetey start --audio-only Record audio only, don't askChoose a window, not just an app. After you opt in, Claude lists the app's windows by title — Weekly Sync — Google Meet rather than "all of Chrome" — and capturing one window keeps the app's other windows and its notification banners out of frame entirely. If more than one display is attached you'll be asked which; otherwise Meetey picks the one showing the meeting.
ScreenCaptureKit has no concept of a browser tab, so a window is as fine as macOS allows. To isolate a single tab, drag it into its own window and pick that. Switching tabs inside the captured window captures the new tab.
Instead of recording video, Meetey samples the screen once a second and keeps a JPEG only when the picture materially changes. An hour-long meeting typically yields a few dozen keyframes rather than 3,600, which is what makes this cheap enough to be worth doing:
Storage/hr | Sent to Claude | |
Audio | ~115 MB | ~12K tokens (transcript) |
Screen keyframes | ~10 MB | ~5K tokens (recognized text) |
Text is recognized on-device with the macOS Vision framework, so the default path sends only text. Claude reads an actual keyframe image only when the text isn't enough — a diagram, a chart, a UI screenshot.
⚠️ Screen capture records everything the target app displays — other tabs, other windows of that app, and notification banners that appear over it. A screen share can contain credentials, customer data, or a colleague's private dashboard. Close or mute anything sensitive before you opt in.
Keyframes are written to ~/.meetey/recordings/<session>-frames/ and are not deleted automatically — the notes reference them, so removing them would make the summary unverifiable. Delete that folder yourself when you're done with it.
Two rules do most of the work of keeping that number small:
A keyframe is written when the screen stops moving, not when it starts. A slide transition collapses to a single file — the settled one, which also reads far better than anything caught mid-fade.
Regions that are always moving stop counting. A speaker tile, a progress bar, a clock: if a part of the screen changes constantly, it is excluded from the comparison. That is what makes the usual slide-plus-speaker-tile layout produce about one keyframe per slide.
Returning to a slide you already showed doesn't write it twice — the manifest records that it came back, and the notes can cite both moments from one image.
Tuning, if you need it (pass through to meetey-capture):
Flag | Default | Purpose |
|
| Frames sampled per second |
|
| Grid cells (of 1024) that must change to count as a new scene. Raise it if you're still getting near-duplicates; lower it if a subtle slide edit was missed |
|
| Hard cap per session |
|
| Force a keyframe if the screen never comes to rest, so a video demo still records something |
| off | Stop ignoring constantly-moving regions |
| off | Skip on-device text recognition |
sceneThreshold, windowID, and displayID are all reachable from /meetey start — just ask.
Camera-heavy calls used to be a poor fit. A moving video tile once produced a keyframe every couple of seconds until the 200-frame cap was hit, usually within the first ten minutes, leaving the rest of the meeting uncaptured. Moving tiles are now ignored, so that no longer happens. A faces-only call still won't produce anything worth reading — stay on audio for those — but it will no longer exhaust the budget and go dark.
Recording without remembering to
Meetey can watch for meetings and offer to record them, so you don't have to think about /meetey start at the moment a call begins.
/meetey watch on Turn it on
/meetey watch Is it on?
/meetey watch off Turn it off
/meetey watch log What it has noticedOr from a terminal: npx jej2k5/meetey watch enable | disable | status | logs.
It never records on its own. When it notices a meeting it asks, naming the window:
┌──────────────────────────────────────────────┐
│ Meetey noticed a meeting: │
│ │
│ Weekly Sync — Google Meet │
│ │
│ Record it? "Audio + screen" captures only │
│ this window. It stops on its own when the │
│ call ends. │
│ │
│ [ Skip ] [ Audio + screen ] [ Audio only ]│
└──────────────────────────────────────────────┘Skip and it won't ask again for that meeting. A prompt you never answer times out as a no, and Return is wired to Skip so an absent-minded keystroke can't start a recording.
Three things worth being clear about:
It is off until you turn it on, and turning it on installs a login agent that persists across restarts until you turn it off.
You choose what it captures. The prompt offers audio only or audio + screen, and screen capture covers just the meeting window it found — narrower than what
/meetey startcaptures by default.It does not detect that a call ended. The recording stops when the app quits or the captured window closes — not when a meeting wraps up while the app stays open. Detection is deliberately loose in the other direction too: it would rather ask about something that isn't a meeting than miss one, since a wrong guess costs one dismissed dialog.
Recordings it makes are transcribed as soon as they end, so they're ready by the time you get to them. /meetey list shows them alongside everything else; /meetey stop works on one that's still running.
If the watcher is on but never asks, check /meetey watch log. could not list windows means Node needs Screen Recording permission — separately from the capture binary — in System Settings → Privacy & Security.
Managing your meetings
Everything about the library is reachable from Claude Code — there's no separate app, dashboard, or web UI to open. Every subcommand has a plain-language equivalent, so use whichever you'd reach for:
Subcommand | Or just ask | Tool |
| "what meetings do I have this week?" |
|
| "show me the standup from Thursday" |
|
| "what did we decide about the API cutover?" |
|
| "delete yesterday's test recording but keep the notes" |
|
| "is meetey working? screen recording seems broken" |
|
/meetey status is about the recording running right now; /meetey doctor is about whether the install itself is healthy.
Search is the one worth knowing about. Asking "what did we decide about the API cutover?" returns the decision, the meeting, and the timestamp — so you can jump straight to that second in the transcript instead of reading it.
What stays local
Stays on your machine, always | Sent to Claude to write the summary |
Meeting audio (the WAV) | The transcript text |
Screen keyframes (the JPEGs) | Recognized on-screen text (OCR) |
Whisper transcription — runs locally | A few keyframe images, only when text isn't enough |
Vision OCR — runs on-device |
Nothing is uploaded to a third-party meeting service, and no bot joins your call. Audio and images never leave your machine unless Claude needs a specific keyframe to interpret a diagram. Summarization itself runs through Claude, so the text does go to the API — the same as anything else you'd paste into Claude Code.
Nothing records without a person saying so. /meetey start is explicit, and the optional watcher asks before every recording — there is no mode in which Meetey captures a meeting you weren't asked about. Screen capture is a separate opt-in on top of that, requested per recording, and the watcher never requests it.
CLI
npx jej2k5/meetey install # First-time setup
npx jej2k5/meetey update # Rebuild binary and refresh MCP server + skill
npx jej2k5/meetey status # Show what's installed and whether everything is wired up
npx jej2k5/meetey watch enable # Watch for meetings and offer to record them
npx jej2k5/meetey watch disable # Stop watching
npx jej2k5/meetey watch status # Is the watcher running?
npx jej2k5/meetey watch logs # Recent watcher activitystatus output:
Meetey status
✔ meetey-capture binary (~/.meetey/meetey-capture/.build/release/meetey-capture)
✔ binary code-signed
✔ whisper-cli (/opt/homebrew/bin/whisper-cli)
✔ Whisper model (ggml-base.en.bin) (~/.meetey/models/ggml-base.en.bin)
○ meeting watcher (optional) (off — enable: meetey watch enable)
✔ MCP server files (~/.meetey/mcp-server)
✔ MCP server registered in ~/.claude.json
✔ /meetey skill installed
✔ Claude Code shortcuts (Ctrl+Shift+R / Ctrl+Shift+S, while Claude Code is focused)Requirements
macOS 13 (Ventura) or later
Node.js 18+
Xcode Command Line Tools (
xcode-select --install)Claude Code
Homebrew (installed automatically if missing)
Whisper model
The default model is ggml-base.en.bin (~142 MB), stored at ~/.meetey/models/. It handles clear English speech well. For better accuracy on accented speech or technical vocabulary, switch to a larger model:
Model | Size | Notes |
| 142 MB | Default. Fast, good for clear English. |
| 466 MB | Better accuracy. |
| 1.1 GB | Best accuracy. Slower. |
To switch, download the model to ~/.meetey/models/ and set MEETEY_MODEL in the mcpServers.meetey.env section of ~/.claude.json.
Troubleshooting
Problem | Fix |
| Run |
| Run |
App not listed by | Ensure the meeting app is open and a call is active |
Blank or garbled transcript | Check Screen Recording permission: System Settings → Privacy & Security → Screen Recording |
Chrome captures wrong audio | Mute other tabs playing audio before starting |
MCP tools not available after install | Restart Claude Code — MCP servers connect at session startup |
| It's a Claude Code shortcut, not a system-wide hotkey — it only fires when Claude Code is focused. Stop from the menu bar instead |
No screen content in the notes | Screen capture is opt-in per meeting — start with |
Too many near-identical keyframes | Narrow to one window when you start; failing that, raise |
Slide changes missed | Lower |
Keyframe text is garbled | Confirm the meeting window isn't scaled down; OCR runs on the captured resolution |
Keyframes from the wrong screen | Pass a |
Recording stopped early | It ends when your call ends, the app quits, or the captured window closes. Rejoining in a new window won't resume it |
Recording didn't stop when the call ended | Only fires if the app took the microphone in the first place — joining by phone while sharing your screen leaves nothing to detect. It falls back to stopping when the app quits |
| If it already stopped itself, |
Watcher never asks |
|
Watcher asks about things that aren't meetings | Expected — it errs toward asking. Say "Not now" and it drops that window |
How it works
/meetey skill → MCP server (Node.js, ~/.meetey/mcp-server/) → meetey-capture (Swift, ~/.meetey/)
watch agent ↗ (~/.meetey/daemon/, optional) → whisper-climeetey-capture is a Swift CLI using ScreenCaptureKit. It takes --app <bundle-id> and --output <path.wav>, records app audio as 16-bit PCM WAV at 16 kHz mono, and stops on SIGTERM, on --stop-after, or via --auto-stop when the app quits or the captured window closes. With --video it also samples the screen, keeps a JPEG when the picture settles into something new, and runs Vision OCR on each one.
MCP server manages the capture process lifecycle and shells out to whisper-cli for transcription. Registered globally in ~/.claude.json so it's available in every Claude Code session.
/meetey skill drives the user-facing flow: list_apps → list_windows → start_recording → stop_recording → transcribe → get_keyframes, then formats and saves the output.
Watch agent is an optional launchd LaunchAgent that notices meetings and asks whether to record them. It is off until enabled and starts nothing without a confirmation. Because it's a separate process from the MCP server, both coordinate through a state file keyed on process liveness — which is why /meetey stop can stop a recording the watcher started.
Two self-tests run without Screen Recording permission or a live meeting:
meetey-capture --selftest # scene detection, settling, dedup, OCR, manifest, auto-stop
node ~/.meetey/daemon/watch.js --selftest # meeting detection patternsFiles written
Path | Contents |
| Raw audio from each session |
| Structured notes — summary, timestamped decisions and action items, screen content |
| Full transcript, one timestamped line per utterance |
| Screen keyframes + |
| Whisper model |
| MCP server (stable install location) |
| Swift binary and sources |
| Watch agent (dormant unless enabled) |
| Which recording is running, if any — removed when it ends |
| Watcher activity (only when enabled) |
| Watcher login agent (only when enabled) |
Contributing
Issues and pull requests are welcome. To work on Meetey from a local clone:
git clone https://github.com/jej2k5/meetey.git
cd meetey
node cli/index.js install # build, sign, and register everything locallyAfter changing Swift sources, rebuild and re-sign:
cd meetey-capture
swift build -c release
codesign --force --sign - --options runtime --entitlements entitlements.plist \
.build/release/meetey-captureThen restart Claude Code so the MCP server reconnects.
New source files should carry the Apache 2.0 header used by the existing files. By contributing, you agree that your contributions are licensed under the Apache License 2.0.
License
Copyright 2026 John Joseph
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the LICENSE file for the specific language governing permissions and limitations under the License.
Third-party attributions are listed in NOTICE.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- Alicense-qualityCmaintenanceCaptures and organizes structured markdown notes from Claude Code sessions with automatic pattern detection, complexity analysis, advanced search, and weekly reports. Everything runs locally without external dependencies.Last updated1MIT
- Alicense-qualityDmaintenanceEnables bidirectional voice interaction for Claude Code using local speech-to-text and text-to-speech models optimized for Apple Silicon. It provides tools to listen to user speech via microphone and speak responses aloud through system speakers.Last updated16Apache 2.0
- AlicenseAqualityAmaintenanceEnables transcription, summarization, and action item extraction from audio files on your Mac using MacWhisper and Claude Desktop, all locally without any cloud APIs.Last updated72MIT
- AlicenseAqualityDmaintenanceLocal speech-to-text transcription using Microsoft's VibeVoice-ASR model with speaker diarization, enabling audio transcription directly in AI tools like Claude Code, Cursor, and OpenCode.Last updated32MIT
Related MCP Connectors
Connect Claude to Fathom meeting recordings, transcripts, and summaries
WHOOP recovery, strain, sleep and workouts in Claude via official WHOOP OAuth. Free, open source.
Search recordings, summarize meetings, create clips, and automate workflows from your AI assistant.
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/jej2k5/meetey'
If you have feedback or need assistance with the MCP directory API, please join our Discord server