AgentCall join-meeting Skill
Gives AI coding agents a voice, an avatar and a screenshare so they can join and take part in video meetings.
- Source repo
- pattern-ai-labs/agentcall
- Stars
- ★ 157
- Last updated
- 8d ago
- License
- MIT
- Primary language
- Python
- FA score
- 46/100 · Major gaps
At a glance
- Works with
- Portable with changesCodex · Claude Code
- You'll need
- Typical use
- A development team running standup in Google Meet lets a coding agent report progress and, in the same breath, inspect the repo, run commands and push a fix.
- Main limitation
- Core function depends on the hosted AgentCall service and an API key; there is no documented self-hosted path, so the service going down or pricing changing directly affects you.
- Source review
- 46/100 · Major gaps
What does this agent do, and when should you use it?
join-meeting is an AgentCall skill that runs on top of an existing coding agent — Claude Code, Cursor, OpenAI Codex CLI, Gemini CLI, Windsurf, Junie, OpenClaw, GitHub Copilot and others — rather than being a standalone app. It lets that agent join Google Meet, Zoom and Microsoft Teams calls as a bot with voice conversation, an animated avatar camera feed, screenshare, real-time transcription and meeting chat. The skill needs Python 3.10+ or Node.js 18+ (aiohttp/websockets or ws) and talks to the AgentCall service over WebSocket using an API key that can be created for you on first run and is persisted to ~/.agentcall/config.json. Its distinguishing claim is that the agent keeps its full session context during the call, so it can search code, edit files, run commands and commit while it talks — the meeting is just another I/O channel.
Given a meeting URL, the skill launches a bot in one of three modes: audio (voice only), webpage-av (voice plus an animated avatar as the camera feed) or webpage-av-screenshare (voice, avatar and live screensharing of URLs, dashboards or slides). Speech output uses text-to-speech with 54 voices across 9 languages and a claimed sub-second latency; on the listening side the agent receives real-time transcripts of what participants say, and it can take screenshots of shared screens at any moment. Two voice strategies are offered: direct, where the agent controls TTS itself, and collaborative, where GetSun voice intelligence manages turn-taking for multi-party conversations. The skill also implements VAD gap buffering, barge-in prevention, auto-interruption detection, sentence-level tracking of what was interrupted, WebSocket reconnection with call-status checks, and crash recovery that rejoins an active call after the agent restarts. Meeting chat can be used to send URLs, code or text that is awkward to speak, and seven built-in templates (pattern, orb, ring, avatar, dashboard, blank, voice-agent) control what the bot displays.
- A development team running standup in Google Meet lets a coding agent report progress and, in the same breath, inspect the repo, run commands and push a fix.
- A support engineer drops the bot into a Zoom call so it can speak scripted answers and paste links or code into the meeting chat.
- Someone who needs meeting notes deploys an audio-mode bot that receives live transcripts and turns them into written records.
- A presenter uses webpage-av-screenshare so the agent can walk an audience through a dashboard, slide deck or web page while sharing its screen.
- In a multi-party call, collaborative voice strategy hands turn-taking to GetSun so the bot does not talk over people.
- When the network blips or the agent process restarts mid-call, reconnection and crash recovery bring it back into the ongoing meeting.
- Teams experimenting with Microsoft Teams (marked Beta) can run the same workflow there.
How do you install or deploy this agent?
Prerequisites: one of the supported coding agents, Python 3.10+ or Node.js 18+, and dependencies installed with pip install aiohttp websockets (Python) or npm install ws (Node.js). Then pick the install path for your agent. Claude Code: /plugin marketplace add pattern-ai-labs/agentcall followed by /plugin install join-meeting@agentcall. OpenClaw: openclaw skills install join-meeting. Cursor 2.5+: /add-plugin in the editor or browse cursor.com/marketplace. OpenAI Codex CLI: drop the skill into ~/.codex/skills/ or run $skill-installer. Gemini CLI: run gemini extensions install https://github.com/pattern-ai-labs/agentcall from your terminal. Junie: npx skills add pattern-ai-labs/agentcall --agent junie. Windsurf: place the folder in .windsurf/skills/. Single-session Claude Code: place it in .claude/skills/. For anything else, git clone https://github.com/pattern-ai-labs/agentcall.git and point the agent at agentcall/SKILL.md. On first run without a key the skill registers a free account via a one-time email code, or you can paste an existing key from https://app.agentcall.dev/api-keys; new accounts include free trial credits.
How do you use this agent?
Tell your agent: “Join this meeting: https://meet.google.com/abc-def-ghi”. Choose a mode (audio, webpage-av, webpage-av-screenshare) and a voice strategy (direct for one-on-one and support calls, collaborative for group meetings). The API key is asked once and stored in ~/.agentcall/config.json. SKILL.md documents modes, events, commands, templates and Claude Code integration; examples/ contains seven working examples; references/ covers the API plus collaborative mode, interruption handling, crash recovery and screenshare.
What are this agent's strengths and limitations?
- Covers three meeting platforms (Google Meet, Zoom, Microsoft Teams with Teams marked Beta) with three run modes and two voice strategies, so one skill spans note-taking through live presenting.
- The agent keeps full coding-session context during the call, so it can search code, edit files, run commands and commit while participating in the meeting.
- The concurrency plumbing is explicit: VAD gap buffering, barge-in prevention, auto-interruption detection, sentence tracking, WebSocket reconnection and crash recovery are all documented features.
- Distributed as an open Agent Skills SKILL.md with install paths for Claude Code, Cursor, Codex CLI, Gemini CLI, Windsurf, Junie, OpenClaw and GitHub Copilot, under a permissive MIT license.
- Core function depends on the hosted AgentCall service and an API key; there is no documented self-hosted path, so the service going down or pricing changing directly affects you.
- Not a standalone app — it requires a supported coding agent plus Python 3.10+ or Node.js 18+ and third-party dependencies (aiohttp/websockets or ws).
- Framework support is uneven: Aider is only partial (no subprocess) and Cursor needs tmux for a PTY, so some environments require extra setup.
- Microsoft Teams support is explicitly labelled Beta, so reliability there is unproven.
- Usage is metered by meeting time: the base plan covers 6 hours with 1 concurrent call, and paid usage is billed per minute.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|
| AgentCall join-meeting Skill This agent | 46 · Major gaps | ★ 157 | 8d ago | Python | Codex · Claude Code |
| CyberVerse Digital Human Assistant | 33 · Major gaps | ★ 1.7k | 13d ago | Python | — |
| Iva Personal Assistant | 85 · Good | ★ 217 | 1d ago | TypeScript | ChatGPT · Codex |
| Qwen Audio Agent | 59 · Major gaps | ★ 2.7k | today | JavaScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
README states the skill joins meetings, speaks, screenshots, and shares screens while retaining full session context, but provides no least-privilege design, user confirmation flow, data-flow description, or sensitive-data policy; API key persistence to ~/.agentcall/config.json is mentioned in one line with no encryption or permission detail; dependencies are listed by name only (aiohttp/websockets/ws) with no version pinning or security audit; external effects (speaking, screenshots, screenshare) have no rollback or revocation mechanism; attribution points only to AgentCall/Pattern AI Labs with an unverified publisher. Evidence is thin, so all trust criteria score 1.
Tests show the bridge returns command.error for unknown commands instead of silently dropping, acks precede domain errors, and failure messages are specific, so self_consistency and failure_messages score 2; dependency availability is only package names in README with no version constraints, lockfile, or offline/degradation notes, scoring 1.
README lists 30+ agent frameworks, three modes, two voice strategies, and use cases, so audience_and_scenarios and environment_fit score 2; trigger_precision scores 2 because tests cover command alias resolution and unknown-command errors; capability_boundaries only vaguely notes Aider partial support and omits meeting-platform limits, concurrency caps, or permission boundaries, scoring 1.
README is well structured with per-platform install steps, so information_architecture and install_notes score 2; naming_stability scores 1 because commands mix type/command aliases with no versioned interface promise; examples_and_faq scores 1 as the examples directory is only referenced, not evidenced; known_limitations scores 1 for the single Aider note; license scores 3 for a complete MIT text; versioning_changelog scores 0 with no CHANGELOG or version number; maintenance_responsibility scores 1 with only an AgentCall byline and no SLA or contact.
output_usability scores 2 with tests evidencing command ack/error behavior; marginal_value scores 2 as voice meeting participation is a differentiated capability; cost_benefit scores 1 because pricing is only a paragraph (6-hour base, per-minute billing) with no comparison to alternatives or resource-consumption detail.
README claims 54 voices, 9 languages, <1s latency, and 7 templates with no benchmarks or source links, so claim_traceability scores 1; tests and README partially corroborate the command set but no external independent source exists, so cross_source_corroboration scores 1; docs do not separate fact from inference, so fact_inference_separation scores 1.
- Publisher identity is unverified; repository ownership and maintenance commitments cannot be independently confirmed.
- API key persistence to a local config file lacks encryption, permission, or rotation details.
- Dependencies are listed by name only with no version pinning or security audit, creating supply-chain uncertainty.
- External effects such as speaking, screenshots, and screenshare have no documented rollback or revocation mechanism.
- Performance and capability figures in README (54 voices, <1s latency, etc.) have no benchmarks or source support.
- No CHANGELOG or version number, leaving interface stability and update path unclear.