Skip to main content

CSGClaw: quick start

← Back to CSGClaw overview

Quick start​

csgclaw onboard
csgclaw serve

The CLI prints a URL (for example http://127.0.0.1:18080/). Open it in a browser to enter the IM workspace.

First launch: CSGClaw tries to auto-detect the Manager agent profile from CSGHub Lite, Codex, or Claude Code. If detection fails, the Web UI still works and opens the Manager profile panel so you can pick a provider and model.

After startup, the terminal usually shows where CSGClaw IM is available and prompts you to open it in a browser (your output may differ slightly):

Example: terminal output after csgclaw serve showing the IM URL

Optional: local inference with CSGHub Lite​

If you prefer to start a local OpenAI-compatible server first, then point csgclaw onboard at it:

  1. Start CSGHub Lite (example):
csghub-lite run Qwen/Qwen3-0.6B-GGUF
  1. Initialize config (interactive; can target http://127.0.0.1:11435/v1 and import models from /v1/models):
csgclaw onboard
  1. Start the local HTTP server:
csgclaw serve
  1. Open the URL printed in the terminal for the built-in IM workspace.

Run the server (foreground / daemon)​

Foreground (keeps the terminal attached):

csgclaw serve

Background:

csgclaw serve --daemon
  • Default log: ~/.csgclaw/server.log
  • Default PID: ~/.csgclaw/server.pid
  • Stop: csgclaw stop

Non-interactive / scripts​

csgclaw onboard --provider csghub-lite --models Qwen/Qwen3-0.6B-GGUF

Custom upstream:

csgclaw onboard --base-url <url> --api-key <key> --models <model[,model...]> [--reasoning-effort <effort>]

More flags: CLI.

Where config is written​

csgclaw onboard writes local config (default ~/.csgclaw/config.toml) with server, bootstrap, sandbox, and channel settings. Model/provider profiles for agents live in agent state and are managed in the Web UI. Sandbox options, Worker overrides, and profile details: repo docs/config.md; field reference on this site: API & config.

Continue with Web UI or CLI.