captain-claw

Captain is a self‑hosted platform that lets you run a fleet of AI agents on your own infrastructure. You install the software on a server you control, choose any large language model, including fully local ones, and avoid per‑seat licensing because the code is MIT‑licensed. Every interaction the agents have with external tools is written to a tamper‑evident log, so you can later verify exactly what was executed.

The system is built for organizations that need traceability and policy compliance when AI generates work. Each team member logs in with their own Google or Microsoft account, giving the agent a distinct identity. When an agent receives a request, such as “draft a reply to the client”, it creates a draft in Gmail or Outlook but does not send it automatically. Before any send operation, the guard logic checks recent drafts and sent messages; if a similar email already exists, the agent refuses to send and flags the action for your review. This prevents the “sent six times in a row” problem described in the listing. The same guard also blocks autonomous work that would create unsolicited drafts, requiring a human to approve each output.

You interact with Captain through a web‑based taskboard that aggregates work from email, WhatsApp, Telegram and scheduled jobs into columns like “Needs me”, “Waiting on others”, “Agents working” and “Done”. A kiosk provides skill cards that launch predefined workflows, and a library stores every file the agents produce, making it easy to retrieve past outputs. For onboarding, you can upload up to twelve reference documents and a set of instructions; the platform then runs an exam of 10 to 120 questions against the agent's real chat, grading each answer and explaining why any mistake occurred. This exam helps you confirm that the agent follows your internal handbook before it handles live tasks.

Typical inputs include a text prompt, attached policy documents, or a contract you paste into the interface. Outputs are usually drafts of emails, structured data updates (such as tracker entries), or status records that appear on the taskboard. The platform records each tool call, whether it queried a document, wrote a draft, or attempted to send an email, so you can audit the exact sequence of actions.