joinly
CLAUDE.md568 starsChanged 16 months agoMITView on GitHub
joinly is a Python middleware that enables AI agents to join and participate in video meetings (Google Meet, Zoom, Teams). It exposes a FastMCP server providing meeting tools (join, leave, speak, transcribe, chat, snapshot, screen share) that any MCP-compatible AI client can use. This is a uv workspace with three packages: Workspace sources are linked locally via [tool.uv.sources]. Each sub-package has its own pyproject.toml and is versioned/released independently (tags: v*, client-v*, common-v*). AudioReader → format conversion → VAD.stream() → utterance…
The file
We split this CLAUDE.md into the rules it gives an agent
2 rules. About 0 are common, 0 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 60
Protocol-based DI: All major components (STT, TTS, VAD, MeetingProvider, controllers) are defined as Protocol classes in joinly/core.py.
- Very specific: line 63
MCP as the public API: All meeting capabilities are MCP tools/resources defined in joinly/server.py.
≈ “Common” is an estimate: at least 1 in 5 of the 26,980 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
CLAUDE.md
This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
Project Overview
joinly is a Python middleware that enables AI agents to join and participate in video meetings (Google Meet, Zoom, Teams). It exposes a FastMCP server providing meeting tools (join, leave, speak, transcribe, chat, snapshot, screen share) that any MCP-compatible AI client can use.
Build & Development Commands
# Install all dependencies (run from repo root)
uv sync --frozen
# Download required ML models (Silero VAD, Whisper, Kokoro TTS)
uv run scripts/download_assets.py
# Lint (ruff checks all rules by default)
uv run ruff check .
uv run ruff check --fix . # autofix
# Format
uv run ruff format .
# Type check
uv run pyright
# Run tests (skips manual tests by default)
uv run pytest
# Run a single test
uv run pytest tests/test_meeting_transcription.py::TestTranscription::test_mcp_transcription -v
# Run manual/e2e tests (requires JOINLY_TEST_MEETING_URL env var)
uv run pytest -m manual
# Start as MCP server
uv run joinly --port 8000
# Start as client (built-in agent joins a meeting directly)
uv run joinly --client <MeetingURL>Workspace Structure
This is a uv workspace with three packages:
| Package | Directory | PyPI name | Purpose |
joinly | joinly/ | joinly | Main MCP server + meeting automation |
joinly-client | client/joinly_client/ | joinly-client | Python client library + LLM conversational agent |
joinly-common | common/joinly_common/ | joinly-common | Shared Pydantic types used by both |
Workspace sources are linked locally via [tool.uv.sources]. Each sub-package has its own pyproject.toml and is versioned/released independently (tags: v*, client-v*, common-v*).
Architecture
Core Design Patterns
- Protocol-based DI: All major components (
STT,TTS,VAD,MeetingProvider, controllers) are defined asProtocolclasses injoinly/core.py.SessionContainer(joinly/container.py) resolves short string tokens (e.g."whisper") to implementations by convention (joinly.services.stt.whisper.WhisperSTT). - ContextVar per-session state:
SettingsandUsagelive inContextVarso each MCP client connection gets isolated configuration. Settings can be overridden per-connection via thejoinly-settingsHTTP header. - EventBus pub/sub: Two event types (
"segment","utterance") injoinly/utils/events.pyloosely couple the transcription pipeline to MCP resource subscriptions. - MCP as the public API: All meeting capabilities are MCP tools/resources defined in
joinly/server.py. The client package connects viaStreamableHttpTransportor directly to aFastMCPinstance.
Audio Pipeline
AudioReader → format conversion → VAD.stream() → utterance boundary detection → STT.stream() → Transcript. The no_speech_event (asyncio.Event) flows from TranscriptionController to SpeechController to enable barge-in/interruption.
Key Modules
joinly/server.py— MCP tool/resource definitions, health endpoint, session lifespanjoinly/session.py—MeetingSessionorchestrates provider + controllersjoinly/container.py— DI container, buildsMeetingSessionfromSettingsjoinly/settings.py—Settings(pydantic-settings,JOINLY_env prefix)joinly/controllers/—DefaultTranscriptionController(VAD→STT pipeline),DefaultSpeechController(text chunking→TTS→audio output with interruption support)joinly/services/— STT (whisper,deepgram,google), TTS (kokoro,elevenlabs,deepgram,google,resemble), VAD (silero,webrtc,hybrid)joinly/providers/browser/— Virtual AV stack: PulseAudio, Xvfb, Playwright Chromium. Platform controllers for Google Meet, Zoom, Teams match URLs viaurl_patternregex.client/joinly_client/agent.py—ConversationalToolAgentbuilt onpydantic-aimodel_request, manages rolling message history and parallel tool executionclient/joinly_client/prompts.py— System prompt templates (dyadic vs multi-party)common/joinly_common/types.py—TranscriptSegment,Transcript,VideoSnapshot,MeetingParticipant,Usage, etc.
Code Style & Conventions
- Ruff:
select = ["ALL"]with specific ignores. Google docstring convention. Line length 88. - Pyright:
typeCheckingMode = "standard"across all packages. - Pre-commit hooks:
ruff(lint+fix),ruff-format,uv-lock(keeps lockfile in sync). - Tests: pytest with
asyncio_mode = "auto", session-scoped event loop. Tests markedmanualrequire a real meeting URL. Integration tests use a mockup browser meeting serving pre-recorded audio. - Commit style: Conventional commits (e.g.
feat(client):,fix:,refactor:,test:).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

