AgentCall Join-Meeting Skill
Let your AI coding agent join Google Meet, Teams, and Zoom meetings with voice, an animated avatar, and live screenshare — while it keeps working.
SKILL.md transparently discloses that all meeting audio/transcripts transit a third-party cloud (agentcall.dev/FirstCall) and warns that any participant can steer the agent, recommending framework-level permission scoping — good data-flow disclosure. API key stored in ~/.agentcall/config.; tunnels use per-call credentials distinct from the API key. Deducted for: agent can autonomously self-register with the user's email and mint a key without per-action confirmation, webpage modes auto-start mic capture on load, and no stated local exit/data-deletion path for cloud-processed meeting content.
Internally consistent docs: parameter table, modes, and command semantics align; timeouts, reconnection and crash recovery are described, and the repo ships pytest tests for the bridge stdin path with explicit command.error reporting instead of silent drops. Deducted for: no CI workflow evidence in a static read; key paths (actual meeting join, GetSun collaborative flow) lack executed reproduction; test coverage limited to the stdin bridge layer.
Trigger descriptions are clear ('join a call, attend a meeting'), with a mode-selection table, fit/non-fit guidance, and STT-aware naming guidance; audience is clear. Deducted for: the core function depends entirely on overseas services (api.agentcall.dev plus Google Meet/Teams/Zoom) with no mainland-China reachability statement — likely unusable there; no Chinese-language scenario documentation beyond TTS voice counts.
Has MIT LICENSE, semantic versioning (1.1.15), a detailed CHANGELOG (including Unreleased and external contributor credit), multi-framework install paths, a full parameter table and layered guides. Deducted for: unverified publisher identity, no maintenance/ownership commitment or update SLA, missing known-limitations section (cloud privacy boundary and STT failure rates only appear scattered inline).
Clear value proposition (agents joining meetings with voice/avatar), directly usable command outputs, free trial lowers adoption cost. Deducted for: static review cannot verify real meeting behavior; effectiveness hinges entirely on cloud availability and plan limits; no comparison against local alternatives (e.g., self-hosted Whisper+TTS pipelines); representative outputs unverified.
Substantive pytest suites for both bridges (including error paths and request_id echo) exist and correspond to CHANGELOG entries — auditable primary material. Deducted for: static read without execution; narrow coverage (stdin layer only); no CI run evidence; YouTube demos and GetSun/FirstCall performance claims (<1s latency, 54 voices) are marketing claims without independent corroboration.
- All meeting audio and transcripts transit a third-party cloud (agentcall.dev); assess data-egress risk before using with sensitive meetings.
- The skill can self-register via email and mint an API key without per-action confirmation; monitor creation of ~/.agentcall/config..
- Any meeting participant can steer the agent via transcripts — restrict the agent with your framework's allowlist/plan-mode during calls.
- Core function depends entirely on overseas services; likely unreachable from mainland-China networks.
- Static review only — no scripts were executed; real join behavior and GetSun latency claims are independently unverified.
What does this skill do, and when should you use it?
AgentCall is an Agent Skill that lets coding agents like Claude Code, Cursor, Codex, and 30+ other frameworks join video meetings as a real participant. The bot can talk (TTS), listen (real-time transcription), see (screenshots), show an animated avatar, and screenshare dynamically. It runs on AgentCall's cloud meeting infrastructure; the agent exchanges JSON events and commands over WebSocket stdio. It requires Python 3.10+ or Node.js 18+ and an AgentCall account that the agent can self-register via a one-time email code.
Runs ./scripts/run.sh <meet-url> to launch a meeting bot; receives JSON events on stdout (lifecycle, transcript.final/partial, participant joins/leaves, chat messages, TTS state) and sends JSON commands on stdin (tts.speak, send_chat, raise_hand, leave, screenshot, screenshare.start, etc.). Supports four modes: audio, webpage-audio, webpage-av, and webpage-av-screenshare; voice strategies are direct (agent fully controls TTS with 54 voices) and collaborative (GetSun voice intelligence handles trigger words, barge-in, and interruptions). The agent serves avatar pages and shared content from a local HTTP server through a secure tunnel, so no port exposure is needed. The API key is persisted in ~/.agentcall/config..
- A developer has their Claude Code agent verbally report deployment status, check code, and fix bugs live during a standup
- Customer support or technical assistance where an agent talks 1-on-1 with a user in direct mode
- An agent attends a team meeting as an animated avatar and screenshares slides or dashboards on command
- Teams wanting meeting minutes run the bot in audio mode to transcribe in real time and answer questions
- Showcasing interactive content via webpage.open — the agent serves a clickable form, report, or code diff participants open in their own browsers
What are this skill's strengths and limitations?
- Works with 30+ agent frameworks (Claude Code, Cursor, Codex, Gemini CLI, Copilot, etc.) — only needs subprocess spawning and stdout reading
- Full feature set: voice, transcription, screenshots, chat, avatar, dynamic screenshare, plus 7 built-in UI templates
- In collaborative mode, GetSun handles barge-in, interruptions, and trigger words so the agent never needs sub-second response times
- Tunnel design requires no public internet exposure; crash recovery and WebSocket auto-reconnect built in
- Free trial credits (360 minutes); the agent can self-register via email with no human dashboard steps
- Depends on the closed-source AgentCall cloud service with per-minute billing beyond the base plan (6 hours, 1 concurrent call)
- Bots take 30–90 seconds to join, and calls are capped by plan duration limits — a new call must be proactively created before the limit
- Transcripts arrive as agent input, so any meeting participant can steer the agent; you must configure your framework's permission system for high-trust workflows
- Screenshare pages render in a headless browser — participants cannot click or scroll, and pages must be designed for 1280x720 with polling-based updates
- Zoom and Teams support is labeled Beta (Google Meet is fastest); no test suite is provided, so real-world multi-platform stability lacks independent evidence
How do you install this skill?
Claude Code: run /plugin marketplace add pattern-ai-labs/agentcall, then /plugin install join-meeting@agentcall. Other agents: OpenClaw — openclaw skills install join-meeting; Cursor — /add-plugin; Codex CLI — drop into ~/.codex/skills/; Gemini CLI — gemini extensions install https://github.com/pattern-ai-labs/agentcall; generic — git clone https://github.com/pattern-ai-labs/agentcall.git and place the folder in your agent's skills directory (e.g. .windsurf/skills/ or .claude/skills/). Dependencies: pip install aiohttp websockets, or cd scripts/node && npm install.
How do you use this skill?
On first use the skill guides self-registration: run python scripts/python/register.py send --email [email protected], retrieve the 6-digit code from the inbox (expires in 10 minutes), then run register.py verify — the key is saved to ~/.agentcall/config.. Then simply tell your agent: "Join this meeting: https://meet.google.com/abc-def-ghi". The bot takes 30–90 seconds to join; wait for call.bot_ready and at least one participant.joined event before it speaks.
How does this skill compare with similar options?
The source does not name specific competitors; it positions itself as an Agent Skill running on top of 30+ agent frameworks rather than a standalone meeting-bot app.