poco-ai/poco-claw

A more beautiful and easier-to-use alternative to OpenClaw. It features a nicer Web UI, built-in IM support, a sandboxed runtime and channel-based team collaboration. Under the hood, it is powered by a Claude Code–based agent.

Poco — Your Pocket Coworker

What it is – Poco is an open‑source, self‑hosted AI‑agent platform that lets you run a persistent, “assistant‑like” coworker inside a web UI. It builds on the ideas of OpenClaw but adds a full‑stack stack (Next.js front‑end, FastAPI back‑end) and a sandboxed execution environment so the agent can install packages, edit files, run commands, browse the web, and interact with GitHub without touching your host system.

Key capabilities

  • Server‑and‑channel collaboration – long‑lived workspaces with channels, DMs, reply threads, and the ability to mention a persistent agent. The agent can turn a message into a tracked task, show progress in a drawer, and publish results as shared files.
  • Secure sandbox – each task runs inside an isolated container. You can mount local directories (self‑hosted only) so the agent works directly on your project code.
  • Beyond a chatbot – plan mode, queueing, termination, project‑level defaults (model, preset, repo, mounts), and multi‑file uploads.
  • Rich UI – artifact preview for HTML, PDF, Markdown, images, videos, Xmind, Excalidraw, Drawio, etc.; playback view to replay command I/O and browser sessions; light/dark themes.
  • Agentic experience – native Claude‑style code assistance (slash commands, plan mode), reusable presets for model/tool configuration, extensible MCP & Skills system, built‑in headless browser for autonomous web research, GitHub repo integration, background execution and scheduled triggers.
  • Integrations & deployment – mobile‑friendly, IM connectors (DingTalk, Feishu, Telegram) with push notifications, one‑click Docker deployment, upcoming cloud subscription, multilingual UI.
  • Smart memory – powered by mem0, the agent retains preferences, project context, and past interactions to become more personalized over time.

Typical use cases

  • Automating repetitive development tasks (dependency installation, code generation, refactoring) while keeping a full audit trail.
  • Collaborative research where multiple users and agents share artifacts and continue work across sessions.
  • Building a “virtual teammate” that can browse the web, pull in data, and update a codebase without manual copy‑pasting.
  • Managing long‑running jobs (e.g., data pipelines) that continue after the browser is closed.

Getting started

# Clone the repo and run the provided script
git clone https://github.com/poco-ai/poco-claw.git
cd poco-claw
./scripts/quickstart.sh   # generates config, builds Docker image, starts services

After the script finishes, open http://localhost:3000 in a browser. Detailed deployment steps and troubleshooting are in the Deployment Guide.

License – MIT (per the badge in the README).

Why it matters – Poco bundles a full agent runtime (LLM back‑end, sandboxed execution, UI, memory) into a single Docker‑deployable product, lowering the barrier for teams that want a private, extensible AI coworker rather than relying on hosted chat services.

Related

  • Project
    agent-infra/sandbox

    A unified agent sandbox environment that combines a browser, shell, and IDE tools in one Docker container with a shared filesystem for AI agents.

  • Project
    AIDotNet/NextCoWork

    A native desktop workspace for coding agents that integrates an editor, terminal, browser, and Git into a single environment with a secure agent kernel.

  • Dispatch
    Securing AI Agents with Isolated Docker Sandboxes

    An exploration of agent-sandbox, a tool designed to run AI coding agents within isolated Docker containers to prevent unauthorized system access and ensure execution safety.

  • Dispatch
    Meko: Solving the Multi-Agent Memory and Knowledge Problem

    Explore how Meko provides a unified, agent-native data infrastructure to enable collective memory, shared knowledge, and decision traceability for multi-agent AI systems.

  • Project
    uvwt/agentdock

    AgentDock is an open‑source runtime that lets LLM chat agents (e.g., ChatGPT) safely perform file operations, run commands, use Git, automate browsers, and coordinate work across multiple computers or containers via a secure HTTP‑based MCP endpoint. It provides fine‑grained permission controls, task persistence, and extensible Skills, but does not include its own chat UI or model inference.