Questions people ask first.

An MIT-licensed local crew chief for the coding agents you already use, not another agent product. One terminal selects a worker for each task from Claude Code, Codex CLI, Cursor, and remote models through OpenCode (Grok, DeepSeek, Gemini, GLM, Kimi, Qwen, MiniMax). It weighs expected quality, speed, cost and availability, recovers eligible provider failures, and keeps an inspectable local ledger.

Best fit: builders who already run several agents or mix subscription CLIs with API models, and want one workflow that routes, recovers, chains and inspects those workers, including /oss, /deterministic, mid-turn /btw / /interrupt, and captain why.

Not a fit: you are happy inside one agent (Cursor, Claude Code, OpenCode alone); you want a managed cloud crew with published product evals (Devin Fusion, Amp); you need multi-tenant SaaS or Windows today. See the best-fit section.

No. Type the task. Routine prompts use fast triage and value ranking; harder work can go to a director that chooses a worker or team. Reach for /frontier, /team, /cheap, /oss, /deterministic, /btw, /interrupt or a leg name only when you want to steer. /btw amends a turn already in flight on supported legs; /interrupt stops it and keeps the work.

/oss limits the turn to open-weight models. /deterministic limits it to legs whose serving tuple is green in the Agentic Determinism Index and pins that stack for the run. Both compose with /repeat, /team and other modifiers. See ADI.

Captain records a reset time when the provider reports one and avoids that worker while it cools down. Eligible failures can retry with another available worker. If useful work already exists, Captain keeps it and marks it partial. Conversation context supports the next turn; provider caches and hidden state do not transfer.

Yes, in one line: /grok draft > /cursor review it + /claude red-team it. > sequences stages, + runs legs in parallel. /wf compiles plain English into that form and shows the plan before running. See Workflows.

Different layers. Fusion optimizes a tuned lead/sidekick loop with persistent contexts and published evaluations. Captain crews your Claude, Codex, Cursor, Grok and API workers: per-task routing, provider recovery, gated workflows, /oss / /deterministic pools, mid-turn steering, and captain why. No Fusion adapter and no head-to-head performance claim ship today. Read the comparison.

Six things most alternatives do not combine: (1) native CLI workers from separate vendors in one conversation, (2) per-task value ranking that can skip the director on routine work, (3) mid-turn /btw and /interrupt, (4) named policy pools (/oss, ADI /deterministic), (5) inspectable decisions (why / quota / budget), (6) gated multi-stage workflows plus captain task mcp for other hosts. Details on the USP list.

No. Routine writing and small edits can use a capable economical worker, while difficult reasoning or consequential changes can justify a stronger model and extra review. Captain’s routine ranking applies a minimum expected-quality bar, then weighs cost, observed speed and subscription pressure. Quality scores are estimates; tests and acceptance still matter. See how routing works.

No. It expresses a saving preference for that turn only. For hard limits, set CAPTAIN_MAX_ATTEMPTS and/or CAPTAIN_MAX_COST (see captain budget). Attempt and time caps are enforced. Dollar caps are tracked in admission mode by default. No adapter can bound every internal tool turn of a delegated run. With CAPTAIN_STRICT=1 and a cost cap, Captain rejects legs that cannot report per-turn cost before dispatch. /cheap does not flip those controls on.

No fixed saving is promised. Economical routing and context pruning can reduce waste, while teams, reviews and retries can add tokens. Your existing subscription charges remain; API costs depend on the work and providers. Captain estimates quota pressure from observed limits rather than measuring every account’s remaining allowance.

Yes. Captain is MIT licensed. Source, issues, and pull requests are at github.com/lemma-ventures/captaincode. Contributions use DCO sign-off. See the open source page.

The harness is free. You pay your existing agent subscriptions and provider bills. Captain Code does not sell seats or proxy keys.

Local CLIs: Claude Code, Codex CLI and Cursor. Subscription/API workers: Grok, Grok Max, Codex via OpenCode, GLM, DeepSeek, Gemini, Qwen, Kimi, MiniMax, Step and DeepSeek V4 Flash on Hugging Face, and a free OpenCode Zen slot. Add any OpenRouter, NVIDIA NIM, or Hugging Face model with captain legs add <id> openrouter/…, nim/…, or huggingface/…. See Adding a leg.

No. Use the CLIs and providers you already have. Doctor shows which legs are wired.

No. It runs on your machine. Prompts go to the providers you choose. There is no Captain Code cloud that stores your repos.

OpenCode is an open source coding agent. Captain Code builds on that foundation to coordinate connected vendor CLIs and remote models with task routing, provider recovery and a shared ledger. OpenCode already supports multiple providers and configurable subagents. See the comparison matrix.

Captain focuses on task routing across connected accounts, provider recovery, model workflows and local run history. Pi emphasizes a customizable agent loop; Jido supplies an Elixir agent framework. See the harness integration notes.

captain task mcp exposes bounded task delegation (plan, submit, inspect, events, artifacts, cancel, resume). The same contract is also on the brain’s local /v1/task HTTP API. See the CLI cheat sheet.

Not yet. macOS and Linux only for now.

No. That would breach most providers' terms and is out of scope. Single-user, local credentials only.

Privately to security@lemma.ventures. See the security policy.