Poco

Poco is an AI Agent execution platform and chat collaboration tool designed for Claude Code, positioned as a safer, more polished, and more user-friendly alternative to OpenClaw. All Agent tasks run in isolated Docker containers where dependencies can be installed, files modified, and commands executed without affecting the host environment.

Core Capabilities#

• Secure Sandbox: Each task runs in an independent Docker container, with local directory bind mount support (self-hosted mode only)
• Native Claude Code Experience: Replicates Slash Commands, Plan Mode, AskQuestion
• MCP & Skills: Supports MCP protocol and custom Skills for unlimited extensibility, with a built-in autonomous browser via Playwright MCP
• Preset System: Project-level pre-configuration of models, toolchains, sub-agents, and visual identities
• Artifacts View: Renders HTML, PDF, Markdown, images, videos, Xmind, Excalidraw, Drawio, and more
• Playback View: Replays command I/O, browser sessions, and tool call processes
• IM Integration: DingTalk, Feishu, Telegram with Stream long-connection mode (no public callback needed)
• Smart Memory: Optional mem0-powered memory system (pgvector + Neo4j) for persisting user preferences and project context
• Background Execution: Agent continues running in the cloud after browser is closed
• OAuth Login: Google / GitHub, with Chinese and English language support

Architecture Overview#

Frontend (Next.js, TypeScript)
    ↓ API proxy /api/v1/*
Backend (Python, FastAPI)
    ↓
Executor Manager (Python, FastAPI)
    ↓ docker.sock spawn
Executor Containers (Python, Claude Agent SDK)
    ↓ S3 (RustFS) for artifacts
PostgreSQL (persistence)

Key components include: Next.js frontend (port 3000), FastAPI backend (port 8000), Executor Manager (port 8001), PostgreSQL 16, and RustFS object storage (S3-compatible). Task scheduling uses a polling (2s) + lease (900s) mechanism, with Executors reporting status via callbacks. Executor images come in lite (lightweight) and full (with desktop/browser stack) variants.

Quick Deployment#

git clone https://github.com/poco-ai/poco-claw.git
cd poco-claw
./scripts/quickstart.sh

Requires Docker and an Anthropic API Key (or compatible endpoint). For manual deployment, copy .env.example to .env, fill in ANTROPIC_API_KEY, then run docker compose up -d. Enable Smart Memory with docker compose --profile mem0 up -d.

Unconfirmed Items#

• Cloud subscription pricing and launch timeline
• Specific purpose of C# accounting for 16.2% of the repository
• Exact relationship with OpenClaw/Manus
• Whether Anthropic is officially involved
• Minimum hardware requirements for deployment
• Compatibility with non-Anthropic models