Everything the harness does.
A production-grade, OpenAI-compatible coding-agent harness. These docs cover the surface area — the full README lives on GitHub.
Getting started
Install catcode, log in to a provider, pick a model, run your first prompt.
Providers & login
Umans, OpenAI, Gemini, Anthropic. Simultaneous logins, OAuth via PKCE, and key precedence.
Subagents & intercom
Eight built-in agents, single/parallel/chain runs, and the peer intercom bus.
Safety & approval
The human-in-the-loop gate, three approval modes, firejail sandbox, and workspace confinement.
Architecture
The Rust core, Go TUI, and Next.js web frontend — one JSONL protocol over stdio.
Plugins
Hook points at pre_/post_/lifecycle/pre_turn. A pre_write linter or vision handoff.
TUI command reference
Slash commands in the terminal. Anything that isn't a command is sent as a prompt to the current model.
/login | Log in / switch provider |
/model [N|substr] | List models, or switch |
/approval <mode> | never · destructive · always |
/sessions | Searchable session picker |
/new | Fresh session (keeps the old) |
/compact | Compact context now |
/attach | Attach an image (vision) |
/abort | Stop the running turn |
Protocol reference
The core reads commands from stdin and writes events to stdout — one JSON object per line. Build your own client.
- ›init
- ›login
- ›send
- ›logout
- ›abort
- ›reset
- ›approve
- ›set_approval
- ›ready
- ›authed
- ›thinking
- ›delta
- ›tool_call
- ›approval_request
- ›tool_result
- ›metrics
- ›done
The complete README
Every flag, env var, and config option — on GitHub.