CAFleet CLI
A CLI that runs independent coding-agent sessions in tmux or herdr and links them through durable messages on a shared broker.
- 56 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- Agents run without a per-command permission prompt and write code and project files in parallel.
- The Director types into member panes, and all members share one database without authentication.
Install
- Works in
Manual install
brew install himkt/tap/cafleetThe first README install command. Then `cafleet setup`, rerun after every upgrade.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Agents run without a per-command permission prompt and write code and project files in parallel.
- The Director types into member panes, and all members share one database without authentication.
This is third-party code. Review the repository files before installing.
What it does
CAFleet runs several independent coding-agent sessions in tmux or herdr and coordinates them through durable messages. Your session, the Director, registers a fleet, creates members, assigns them research, implementation and review and receives replies while it keeps working. A member reports ready with its first broker command, then performs the assignment and reports back. The broker is a SQLite access layer inside the CLI itself. The `cafleet setup` command installs embedded skills, presets and plugins for the chosen agent offline. Claude Code, Codex, opencode, Hermes Agent and pi backends are supported. A design doc workflow runs in two requests: Create a design doc for ..., then Implement design-docs/...
Who it is for. Developers running several agents in parallel in a terminal who want the agents to pass tasks and replies to each other.
Good fit when
- Research, implementation and independent review go to different agents and results must be collected.
- A feature goes through a design doc: draft, review, approval, then implementation and checks.
- Team members run on different backends, for example Claude Code and Codex.
- Agent replies must outlive a closed pane and be read later.
Not a fit when
- The agent works outside tmux and herdr: orchestration relies on multiplexer panes.
- Agents may not run without a per-command permission prompt: CAFleet needs that setup.
- A single session without task distribution is enough.
Example request
Create a fleet, assign one member to research a library and another to review my diff, and collect the replies.Limitations
CAFleet runs inside a coding agent without per-command permission prompts, so each backend must be configured: a permission profile in `~/.claude/settings.json` for Claude Code, sandbox entries in `~/.codex/config.toml` for Codex. Claude Code members always start in `auto` mode, which must be available for the account and model. After every upgrade, patch releases included, rerun `cafleet setup`: each version starts from a fresh database. A fleet partitions routing but supplies no authentication, and members share one database URL.
How to disable. Close the fleet with Director commands and delete the `cafleet` binary and the installed skills and plugins. Remove the permissions added for CAFleet from the agent settings.
FAQ
How do I install CAFleet?
`brew install himkt/tap/cafleet`, then `cafleet setup`. Or download the archive for your platform from the releases page, put the single `cafleet` binary on `PATH` and run the same setup.
Which agents are supported?
Claude Code (`claude`), OpenAI Codex CLI (`codex`), `opencode`, Hermes Agent (`hermes`) and the pi coding agent. The backend is chosen with the `--coding-agent` flag during setup.
What are the Director and a member?
The Director is your current session: it creates the fleet, assigns tasks and closes the team. A member is a separate coding-agent session that performs an assignment and reports through the broker.
How does the design doc workflow work?
Two requests: Create a design doc for <description> starts drafting and review, and Implement design-docs/NNNNNNN-<slug> starts implementation, checks and a fresh review. The result lands in the repository `design-docs/`.
What do I do after an upgrade?
Run `cafleet setup` again. Every version, patches included, starts from a fresh database.
Discussion
No questions or reviews yet.
Related
CLI toolsA CLI that plugs third-party models into Claude Code orchestration: workflows, agent teams and subagents on DeepSeek, GLM, Kimi, Qwen or a Codex subscription
A terminal workspace where several CLI agents like Codex, Claude and Gemini exchange messages and work on one task under human oversight
An agent orchestration platform with a 670-tool MCP server: Claude Code, Cursor and Codex drive DAG workflows and agent crews.
A CLI connects installed coding agents into a team, dispatches work between sessions and collects results across machines.
One browser window for Claude Code, Codex, Gemini CLI and OpenCode: session status, history and requests between agents.
A tmux-based TUI that runs Claude Code, Codex and other agents in one list: status, quick prompts, worktrees and diff review
README in short
The README briefly describes the idea and two install options, with details moved to the documentation site. It has a quickstart, concepts, members, a specification and a contributor guide.