Several agents, one workflow
You already pay for more than one subscription or mix CLI agents with cheap API models. You want one conversation that routes, recovers and chains them.
Your agents. One captain.
Captain picks the best model(s).
For the task you prompted.
Type the work. Captain classifies it and picks the model or team from the subscriptions, keys, and open-source models you already have. Frontier only when it earns it.
Who is this for
For builders who already run more than one agent.
Captain is the local crew chief, not another coding agent. It earns its place when choosing, switching, recovering and reviewing across Claude, Codex, Cursor, Grok and API models is the bottleneck. Your providers still run inference and bill for usage.
Spend intelligence on purpose
Routine work should stay lean. Hard work can take frontier, a team, or a gated review. Inspect every pick with captain why.
Policy over the pool
Need open-weight only, ADI-deterministic serving, mid-turn steering or a live director helm? Keep your tools; constrain how Captain picks.
Keep the tools that work. Add the crew layer.
Fusion tunes a lead and sidekick. Cursor owns the editor. Amp owns an integrated agent. OpenCode is the foundation. Captain coordinates your workers across those worlds, with recovery, gated workflows and inspectable decisions.
See the full fit, and how it differs →Shipped differentiators: /oss, /deterministic, /btw, /interrupt, /captain helm, captain why / quota / budget, and captain task mcp. Next: publish the pilot baseline and productize host packaging. Read the roadmap →
Automatic until you want the wheel
Type a task.
Or set the direction.
/frontier
Call on Claude at maximum effort for a demanding task.
/frontier investigate the deadlock/team
Ask the director to assemble several workers and synthesize their answers.
/team /frontier review this design/cheap
Give this turn a cost-saving preference. Use /speed or /quality to shift the priority.
/cheap update the changelog/repeat
Work through a task in rounds, with a limit you choose.
/repeat 5 fix the next failing test/workflow
Describe a pipeline, preview it, then run it. /wf is the short form.
/workflow grok drafts, claude reviews/parallel
Start a separate task alongside your main conversation.
/parallel review the release notes/btw <note>
Steer the running worker mid-turn (supported legs only).
/btw add a null check before the divide/oss
Open-weight models only for this turn.
/oss /repeat 5 fix the flaky test/deterministic
Pick an ADI-green serving tuple and pin it.
/deterministic emit the spawn table from the level brief/interrupt
Stop the running worker and keep its work as a handoff or partial.
/interrupt switch approachYou can also name an agent directly: /claude, /codex, /cursor, /grok or any other configured leg. See supported legs →
Captain picks the best model(s) for the task you prompted
Remove the hassle.
Leverage every subscription to the max.
Save tokens. Save headaches. Captain classifies every prompt and routes it to the right model or team from the subscriptions and keys you already pay for, and from the strongest open-source models you connect. One TUI. Unified memory and ledger. No window hell.
Captain picks so you don’t shop
Routine to economical or open-weight. Hard work to frontier or a team. Your existing subs, APIs and open-source legs inform the choice. No manual model roulette.
Keep moving when limits hit
Quota? Captain records reset times and hands off to another available worker with context intact. Useful partial output is preserved, not lost.
Steer when you need to
Use /frontier /team /repeat /workflow /cheap /parallel or explicit chains with > and +. Bare prompts are the happy path.
Workers you can connect
Built-in legs.
Plus any OpenRouter, NIM, or Hugging Face model.
A leg is one named worker. Captain ships with local subscription CLIs and remote models through OpenCode. Add more without rebuilding.
Claude · Codex · Cursor
/claude, /codex-cli and /cursor drive the vendor CLIs you already authenticate. Codex via OpenCode (/codex) uses the ChatGPT subscription path.
Grok · Grok Max
/grok is a fast coding worker on SuperGrok when that account is available. /grok-max is the frontier-class xAI option for harder reasoning. The director helm picks the first runnable capable leg on your machine.
GLM · Kimi · MiniMax · more
Compiled legs include GLM, DeepSeek, Gemini and Qwen on OpenRouter; Kimi on NVIDIA NIM; Step and DeepSeek V4 Flash on Hugging Face; plus a free OpenCode Zen slot. Add any other hosted model the same way. Hugging Face legs count as open weights for /oss.
Add a model from OpenRouter, NIM, or Hugging Face
captain legs add muse openrouter/meta/muse-spark-1.3 --prior 7.9
captain legs add kimi-flash nim/moonshotai/kimi-k2-thinking --prior 7.5
captain legs add ds-r1 huggingface/deepseek-ai/DeepSeek-R1 --prior 7.0
captain doctor
provider/model · first slash splits the provider
openrouter · prices and context from the public catalog
nim · NVIDIA Integrate API (same as nvidia)
huggingface · router models · join the
/oss pool
Restart the brain and relaunch the terminal after adding a leg. Force it once with /muse …, or leave bare prompts to captain.
Your agents. One captain.
Put your setup to work.
Connect your existing agents and keys. Run doctor to see what is ready. Give Captain the first task.
MIT licensed. macOS and Linux. No Captain subscription; provider charges and limits still apply.