AgentGUI多智能体工作流

System Requirements

• Node.js 18+ (LTS recommended)
• SQLite 3
• Modern browser (Chrome, Firefox, Safari, Edge)
• At least one supported AI coding agent installed (see table above)

Manual Installation

git clone https://github.com/AnEntrypoint/agentgui.git
cd agentgui
npm install
npm run dev

Server starts on http://localhost:3000/gm/

Quick Start

Use Cases

Multi-Agent Comparison: Run the same task through different agents to compare approaches, quality, and speed.

Long-Running Projects: Build complex features across multiple sessions without losing context or conversation history.

Team Collaboration: Share conversation URLs and working directories for pair programming with AI agents.

Agent Development: Test and debug custom agents with full visibility into streaming events and tool calls.

Offline Speech: Use local speech-to-text and text-to-speech without API costs or internet dependency.

Screenshots

Screenshots are regenerated on every push to main by .github/workflows/gh-pages.yml using the committed fixture DB (fixtures/data.db) so the gallery stays in sync with the UI.

Environment Variables

• PORT - Server port (default: 3000)
• BASE_URL - URL prefix (default: /gm)
• STARTUP_CWD - Working directory passed to agents
• HOT_RELOAD - Enable watch mode (default: true)
• DEBUG - Enable internal state inspection endpoints (set to 1)

REST API

All routes prefixed with /gm:

Conversations: - GET /api/conversations - List conversations - POST /api/conversations - Create conversation - GET /api/conversations/:id - Get conversation with streaming status - POST /api/conversations/:id/messages - Send message - DELETE /api/conversations/:id - Delete conversation

Agents & Tools: - GET /api/agents - List discovered agents - GET /api/tools - List detected tools with installation status - POST /api/tools/:id/install - Install tool - POST /api/tools/:id/update - Update tool

Speech: - POST /api/stt - Speech-to-text (raw audio input) - POST /api/tts - Text-to-speech (returns audio) - GET /api/speech-status - Check model download progress

WebSocket: /gm/sync - Subscribe to conversation/session updates with events like streaming_start, streaming_progress, streaming_complete

Debug API

When DEBUG=1 is set, internal state inspection endpoints become available:

Endpoints: - GET /api/debug/machines - Inspect all XState v5 machine snapshots (execution, acp-server, tool-install states) - GET /api/debug/state - Full server state dump (connections, queues, active sessions) - GET /api/debug/ws-stats - WebSocket connection metrics and lag distribution

Browser Console (window.__debug): - window.__debug.machines - All XState machines (conv, toolInstall, voice, convList, prompt, recording, terminal, ws) - window.__debug.ws - WebSocket state, latency EMA, latency trend, URL - window.__debug.auth - Agent auth and OAuth state - window.__debug.perf - Conversation perf metrics - window.__debug.config - Base URL and server config - window.__debug.renderer - Streaming renderer event queue / history lengths - window.__debug.conv - Current conversation and streaming set - window.__debug.getState() - WS latency snapshot - window.__debug.getSyncState() - Flat snapshot of all machines (legacy shape) - window.__debug.getMessageState() - Message/queue DOM counts

See CLAUDE.md for complete XState v5 machine documentation and internal architecture details.

Troubleshooting