GotchiBot

Running more than one autonomous agent quickly stops being a model problem and becomes an operations problem. Each agent needs its own credentials, its own working directory, its own model tier, and its own restart policy, and the moment something goes wrong you need to know which agent did it and how to stop it. Doing that with loose shell scripts does not survive past about three agents.

GotchiBot is a cross-platform orchestrator that treats agents as managed processes with identity. It supports two topologies — Solo, where one operator runs agents on a single machine, and Fleet, where an always-on host coordinates many. A tmux-based cockpit gives a live view of every running agent, with spawn, attach, and monitoring as first-class commands. Agents get scoped credentials through abracadabra rather than environment files, model tiering is declared in configuration rather than hard-coded, and MCP servers expose the orchestrator's own capabilities back to the agents it runs. Onboarding is wallet-based: an operator connects a wallet, receives a per-install token, and authenticates against shared infrastructure without ever holding operator secrets. The fleet also has working tools rather than just a supervisor: a PyMuPDF-backed PDF tool exposed over CLI and MCP so agents can extract, search, and chunk documents for retrieval; a meeting layer that wakes agents on a checklist, chairs their summaries with a question round, and writes end-of-meeting handoff notes; and a bridge that routes fleet prompts through the Claude Code CLI on the host, with jobs submitted asynchronously and collected when the answer lands.

GotchiBot runs a live agent fleet on a home host behind a Cloudflare tunnel, coordinating agent processes, a Docker-backed simulation service, a MongoDB store, and a self-hosted Ollama content pipeline. It ships as open source under PolyForm Noncommercial with a separate commercial entitlement, and installs update themselves from a signed release manifest.