English | 中文

CSGClaw

Your Personal AI Team

CSGClaw is a multi-agent collaboration platform built by OpenCSG — designed around one practical question: once work becomes non-trivial, how do you get a group of AI agents to operate like a team, without the system becoming heavy or painful to set up?

Share a local host agent with other CSGClaw users through an outbound relay connection: see the CSGRelay guide.

Install

macOS / Linux:

curl -fsSL https://csgclaw.opencsg.com/install.sh | bash

Windows (PowerShell):

curl.exe -fsSL https://csgclaw.opencsg.com/install.ps1 | powershell -ExecutionPolicy Bypass -Command -

The installers download a prebuilt release bundle, install it into user-local directories, and put csgclaw on your PATH.

• install.sh currently supports macOS arm64, Linux amd64, and Linux arm64.
• install.ps1 currently supports Windows amd64.

Official release bundles are also published for macOS amd64 and Windows amd64. Local runtime execution does not require Docker.

Build from source:

make build

See docs/tech/build.md for build targets and companion host CLI packaging.

Build the desktop app (Electron):

make desktop-package

This builds a desktop installer for the current platform. Outputs are written to desktop/out/make/.

On Windows without make:

.\scripts\build.cmd desktop-package

For signed and notarized macOS release packages, see the desktop signing and release guide.

For most users, the install script above is the simpler option.

Quick Start

csgclaw serve

CSGClaw prints the IM workspace URL without opening a browser by default. Use csgclaw serve --browser to open it automatically after startup.

Configuration

csgclaw serve uses a local config with server, bootstrap and channel settings, and auto-creates any missing bootstrap state on first run. Agent model/provider profiles are stored in agent state and managed from the Web UI. See docs/tech/config.md for runtime deployment, migration and Agent profile details.

Maintainer-facing architecture, API, structured skill output, and IM thread design notes live in docs/tech/architecture.md, docs/tech/api.md, docs/tech/structured-output.md, and docs/tech/im/im-threads.md.

Features

• Multi-agent coordination — work with a team of specialized agents through a single coordination point, not a pile of chat windows
• One-click install — prebuilt releases for macOS arm64/amd64, Linux amd64/arm64, and Windows amd64; up and running in minutes
• WebUI out of the box — browser-based workspace available immediately after csgclaw serve
• Multi-channel support — connect Feishu, WeChat, Matrix, or other channels when needed
• Runtime access controls - Workers use the configured runtime execution mode and permissions
• Role-based Workers — specialize Workers for frontend, backend, testing, docs, research, and more

What CSGClaw Is

CSGClaw gives you one Manager and a set of specialized Workers, so instead of juggling isolated agents, you work through a single coordination point for defining goals, breaking down work, assigning roles, tracking progress, and collecting results.

┌────────────────────────────────────────────────────────────┐
│                         CSGClaw                            │
│  ┌──────────────────────────────────────────────────────┐  │
│  │ Manager — understands goals, plans, coordinates      │  │
│  └──────────────────────────────────────────────────────┘  │
│               ↓                      ↓                     │
│        Worker Alice            Worker Bob                  │
│          frontend                 backend                  │
│                                                            │
│   WebUI / Feishu / WeChat / Matrix / other channels        │
└────────────────────────────────────────────────────────────┘
                      ↑ you make decisions

Manager — receives your goals, decomposes tasks, selects Workers, tracks progress, and consolidates results.

Workers — role-specific executors (frontend, backend, testing, docs, research…). Specialization keeps context clean and reduces role confusion.

Runtime environment - Codex and DSH run as local processes without Docker or BoxLite.

Interface — WebUI out of the box; Feishu, WeChat, Matrix, and other channels available as integrations.

A Typical Workflow

You: Build a web app prototype — landing page, login, and basic admin view.

Manager: Splitting into tasks.
  · Alice → landing page & login UI
  · Bob   → backend APIs & data model
  · Carol → integration checks

You: Add GitHub login to the login flow.

Manager: Updating Alice and Bob.

Carol: Login response is missing the user avatar field.

Manager: Bob updates the API first; Alice updates the UI once the field contract is confirmed.

The key isn't that multiple agents exist — it's that their collaboration is organized.

Design Principles

Codex and DeepSeek Harness runtimes. CSGClaw uses the bundled Codex CLI for the Manager and supports Codex and DeepSeek Harness workers through a shared Agent Engine.

Local runtime execution. Codex and DSH execute in the environment running CSGClaw, with access controlled by runtime permissions and execution modes.

WebUI first, channel-agnostic by design. Many multi-agent systems are tightly coupled to one messaging protocol. CSGClaw ships with a built-in WebUI so you can start immediately, while keeping other channels (Feishu, WeChat, Matrix) as optional integrations — not assumptions.

Who It Is For

• Independent developers who want an AI team, not just a single assistant
• Small teams that want lower-friction multi-agent collaboration
• Users who value fast startup, lighter runtime, and sensible defaults

Acknowledgement

CSGClaw is informed by ideas explored in HiClaw around multi-agent usability, while placing stronger emphasis on lightweight runtime, easier local startup, and a platform model not bound to a single communication channel.

License

CSGClaw is licensed under the Apache License 2.0. See LICENSE for details.