pattern-ai-labs/agentcall

Purpose and Integration Surface

join-meeting is an AgentCall skill, not a standalone application. It lets coding-agent runtimes join Google Meet, Zoom, or Teams and treat a live meeting as an I/O channel for speech, transcripts, chat, screenshots, avatar video, and screensharing. The primary agent-facing contract is SKILL.md; platform packaging is supplied by .claude-plugin/plugin.json and gemini-extension.json.

Runtime Entrypoints

The skill selects either Python 3.10+ or Node.js 18+ scripts. Shell dispatch is centralized in scripts/run.sh. Equivalent language implementations expose operational commands:

• Join/control call: scripts/python/join.py, scripts/node/join.js

• Bidirectional meeting bridge: scripts/python/bridge.py, scripts/node/bridge.js

• Agent-initiated outbound calls: scripts/python/agentcall.py, scripts/node/agentcall.js

These subprocesses communicate status, events, and commands through stdout/stdin so a host coding agent can supervise them.

Identity, Registration, and Service Boundary

Before calls, registration and API-key handling establish access to AgentCall’s hosted service. scripts/python/register.py and scripts/node/register.js implement first-run account/key setup; keys persist outside the repository in the user AgentCall configuration. scripts/python/check_update.py and scripts/node/check_update.js support lifecycle/version checks. AgentCall is the central external service for meeting transport, TTS, transcription, call state, and associated hosted capabilities.

Meeting Connection and Event Bridge

join creates or reconnects a meeting participant, while bridge maintains the long-running live interaction channel. The bridge receives real-time transcript/call events, accepts instructions such as speaking or chat actions, and manages WebSocket reconnection and active-call recovery. Visual-mode variants, scripts/python/bridge-visual.py and scripts/node/bridge-visual.js, add camera/avatar and presentation-oriented capabilities. The boundary is explicit: agent reasoning remains in the host agent process; media/session operations remain in AgentCall scripts and APIs.

Agent Control Loop

A coding agent launches a bridge subprocess, reads emitted events, decides what to say or do using its existing workspace context, then writes commands back to stdin. scripts/node/stdin-commands.js is the Node command parser/control layer; Python bridge behavior is exercised through tests/test_bridge_stdin.py. This design permits direct agent-controlled speech or collaborative timing through GetSun voice intelligence, without embedding an LLM runtime in this repository.

Visual and Presentation Pipeline

Webpage AV modes use hosted/custom webpages as the bot’s camera feed and can screenshare dynamic URLs. Reusable HTML/JS sources are under ui-templates/, including ui-templates/avatar/index.html, ui-templates/dashboard/index.html, and ui-templates/voice-agent/index.html. Audio-oriented browser behavior is represented by ui-templates/agentcall-audio.js. scripts/python/tunnel.py and scripts/node/tunnel.js expose local content for meeting-accessible presentation or visual URLs.

Operational Patterns and Reference Implementations

The repository documents its protocol and advanced control semantics in references/api.md, references/guides/collaborative-mode.md, and references/guides/crash-recovery.md. Runnable examples demonstrate the intended host-agent orchestration rather than a shared application framework: examples/meeting-assistant/assistant.py, examples/meeting-assistant-smart/assistant.py, and examples/coding-companion/bridge.py. They show how transcript events, voice responses, tools, and persistent coding context combine around the bridge.