claude-office

Office mode for Claude Code: your session model stays the boss (judgment, planning, final review) and delegates execution to a roster of headless agents running under other models — with cross-model-family review, so a model never grades its own biases.

The roster

Agent	Name	Code	Model	Role
boss	—	—	you (session model)	judgment, planning, synthesis, final review
lead	Tanaka	tnk	qwen3.8-max-preview	main executor: features, multi-file edits
ds	Suzuki	suz	deepseek-v4-flash-0731	executor B + adversarial reviewer
glm	Kenji	ken	glm-5.2	executor C + second reviewer
k2	Yuki	yuk	kimi-for-coding	fallback executor (separate credit pool)
flash	Hayate	hay	qwen3.6-flash	mechanical bulk work

Usage

/office                       # activate for the session
/office off                   # deactivate

When active, the session model spawns roster agents via the bundled office-agent script (headless claude -p routed through a local LiteLLM proxy or the Kimi endpoint), fans work out in parallel, and runs cross-family review (author != reviewer) before merging.

Getting started

You don't need this exact roster. Any OpenAI-compatible provider (Alibaba MaaS, DeepSeek, OpenRouter, Together, local Ollama...) works through the bundled LiteLLM setup, and the roster is just a table at the top of office-agent you edit to match.

Full walkthrough: SETUP.md — install, proxy config (examples/litellm-office.yaml), keeping it alive (examples/com.office.litellm.plist), pointing the roster at your models, statusline integration (examples/statusline-office.sh) and troubleshooting.

Quick version:

git clone https://github.com/Wecko-ai/claude-office.git /tmp/claude-office
cp -r /tmp/claude-office ~/.claude/skills/office
cp /tmp/claude-office/office-agent /tmp/claude-office/office-watch ~/bin/
# configure the proxy (SETUP.md step 2), then:
office-agent lead --task "smoke test" "reply with just: pong"

Then /office in any Claude Code session.

Live office view

Two ways to watch the agents work:

• office-hq — localhost dashboard (auto-started by /office, or office-hq --daemon): http://localhost:4545 shows every desk with task + real-time action, sessions grouped, and the recent runs feed. Zero dependencies, stdlib only. office-hq --stop kills it.
• office-watch — same thing as a terminal TUI for a split pane.

Requirements

• claude CLI on PATH (agents are headless claude -p processes) + jq, zsh, python3
• An API key for at least one OpenAI-compatible provider + pip install litellm
• Optional: Kimi coding key in ~/.config/kimi/api_key for k2
• If a backend is missing, its agents are skipped and the boss says so

Full operational manual: SKILL.md.

Install (as plugin)

The skill follows the Claude Code plugin layout (.claude-plugin/plugin.json), so it can also be installed from any marketplace that references this repo.

License

MIT — Wecko