claude-anyteam 0.4.0

Bring Codex and Gemini CLI powered teammates into Claude Code with multi-backend routing.

Native Claude Code teammates, any LLM.

Codex and Gemini today. Kimi, GLM, DeepSeek next — on the same team-native architecture.

Quickstart · Architecture · Roadmap

---

What it is

Claude Code's Agent Teams feature is built for multi-agent collaboration — but every teammate is a Claude instance. claude-anyteam makes it possible for any external model to join the same team, with the same native UX, without wrapping it inside a Claude LLM.

Your Claude Code session orchestrates. External models execute. No chat-wrapper overhead. No "Claude pretending to be Codex." Real models, real teammates.

Quickstart

npx --yes claude-anyteam

That's the entire install. The installer:

• Detects python3 and installs uv if missing (non-interactive, no shell profile edits)
• Installs the claude-anyteam Python tool via uv tool install
• Runs claude-anyteam install (verifies tmux/psmux, probes for the OpenAI Codex CLI and Gemini CLI, warns if either is missing or Codex is below 0.120, writes ~/.claude/settings.json + ~/.claude.json, records install-state for symmetric uninstall)

Restart Claude Code, enable Agent Teams mode, and create a teammate named codex-<anything> or gemini-<anything>:

codex-alice      → routed to claude-anyteam + Codex
codex-reviewer   → routed to claude-anyteam + Codex
gemini-alice     → routed to claude-anyteam + Gemini CLI
gemini-reviewer  → routed to claude-anyteam + Gemini CLI
alice            → native Claude (unchanged)

Codex- and Gemini-prefixed names appear in your TUI presence line exactly like native teammates. Single-terminal mode or tmux — both work.

Why it feels native

Real teammate protocol Not a chat wrapper. The adapter speaks Claude Code's agent-team protocol directly: mailbox I/O, atomic task claims, idle notifications, shutdown lifecycle. A Codex teammate is functionally indistinguishable from a native Claude teammate.	Mid-task reactivity When a peer messages a working teammate, the adapter injects the message mid-turn via Codex's turn/steer App Server call. Codex reshapes the in-flight turn instead of discarding it. v7.1.
Cross-task memory Each new task forks from the previous task's Codex thread via thread/fork. The teammate carries its own conversational context forward across the team's task list. v7.3.	Battle-tested parity 348 passing tests. Ten parity bugs caught by a live 4-teammate hunt (mixed Claude + Codex) and fixed. Zero accepted limitations on the protocol layer.

Supported backends

Backend	Teammate prefix	Status	Notes
Codex via OpenAI Codex CLI 0.120+	codex-*	✅ Supported today	App Server mode for mid-task steer and thread/fork; fresh-exec fallback with codex exec resume.
Gemini via Gemini CLI	gemini-*	✅ Supported today	Default headless gemini --prompt ... --output-format stream-json, plus ACP via gemini-anyteam --backend acp; ACP supports `--trust default

Coming next

Coming next
⏳ Kimi adapter
⏳ GLM adapter
⏳ DeepSeek adapter
⏳ Generic CLI adapter template

Codex and Gemini are shipping. Everything in "coming next" is on the same architectural surface — each new model is a new adapter binary + one line in the spawn shim's routing table. See docs/roadmap.md.

Requirements

• Python 3.12+
• Node 18+ (for the npm installer; not required at runtime)
• OpenAI Codex CLI 0.120+ on PATH for codex-* teammates
• Gemini CLI on PATH for gemini-* teammates
• Claude Code 2.1+ with Agent Teams mode
• Terminal multiplexer on PATH (tmux or psmux) — see configuration.md

Docs

• Install — how the installer wires Claude Code, alternative install methods, headless launches
• Architecture — how the adapter integrates with Claude Code's team protocol
• Roadmap — supported today vs coming next, contribution pointers
• Configuration — CLI flags, env vars, advanced modes
• Releasing — maintainer-facing tag-triggered publish flow

License

MIT