— The real difference A crew isn't just many terminals.
Running four agents in parallel gives you four transcripts to reconcile. Octopush's
Direct mode gives you one pipeline: a cheap model plans, a strong one implements,
another reviews — each stage with its own role, model, substrate (API or CLI) and
reasoning effort, human checkpoint gates where you sign off, review loops, and a
single escalation only after a proven failure. The ledger leads with what you saved.
◪ Per-stage model routing
Each stage gets the model it deserves and its own reasoning effort — plan on Haiku, implement on Opus, review on Sonnet — instead of paying premium for every token.
$ A savings ledger
Every run shows dollars saved against an all-premium baseline. The number nobody else puts on screen: what routing actually saved you.
⟜ Gates & review loops
Human checkpoints pause the crew for your sign-off; review loops send work back with a bounded retry count. You stay editor-in-chief.
⛨ Sandboxed, free
Agents run write-confined under the macOS seatbelt sandbox. Security is never behind the paywall.
◷ Routines & detached runs
Schedule crews (Days × Times, with an optional shell fire-condition) and let them run detached — they survive quitting the app, then notify you.
⇄ Any provider, any gateway
BYOK across Anthropic, OpenAI, DeepSeek and local Ollama — two wire protocols and custom base URLs, so your gateway or local model just works.