Muvon/octomind

Interactive, stdin, daemon, WebSocket, or ACP — one binary, one config. A supervising model plans and checks the work. Policy lives in TOML guardrails, not approval prompts.
Website · Docs · Tap Registry · Benchmarks · Issues
Octomind is an open-source coding agent harness. It gives any LLM the tools to work in a repository — read and edit files, run shells, search code, delegate to sub-agents — through MCP, and runs the result as an interactive CLI, a stdin pipe, a daemon, a WebSocket server, or an ACP agent in your editor. Most agents stop and ask a human before each risky step, which works at a keyboard and blocks everything in CI. Octomind writes that policy down as TOML guardrails and adds a supervising model that plans and checks the work, so the same session can run with nobody watching. On 25 tasks taken from merged pull requests it solved 24; OpenCode solved 19 with the same model at about twice the cost (benchmarks).
## How It Works
One diagram, the whole turn: what is assembled before the model sees the request, the tool loop, how a turn is allowed to end, and what carries over to the next session.
What the model sees is built once per user message, not per tool round: AGENTS.md, the skills whose rules match the prompt (standard SKILL.md folders), the tool bundles those skills need (they start their own MCP servers on demand), and the few past lessons that match. Skills · Learning
Compression folds older turns into a summary that cites content-addressed blocks; the raw messages stay in a per-session archive and the recall tool brings any block back verbatim as a normal tool result, so history is never rewritten and the prompt cache survives. Compression
Guards, hooks, validators are deterministic and live in .agents/guardrails.toml plus scripts; they run before any model opinion. The optional intent check ([supervisor.authorizer], off by default) adds one supervisor-model look after the guards and blocks a call only when it independently conflicts with what you asked for. Guardrails
The supervisor is a second model ([supervisor.model]): it keeps the plan and reminds the model of the goal each turn, cuts oversized tool results down to the lines the task needs (copied, never retyped), and refuses a self-reported "done" until the evidence checks out. The loop check is a fixed heuristic: when the model repeats itself or stalls it gets one nudge, never a takeover. Supervisor
The decision model ([decision]) is optional: a classifier that answers typed yes/no questions with a probability instead of prose. When set, the supervisor sends its cheap questions there — which lessons to recall, which skill to activate, which lines to keep, whether a completion claim holds — and keeps chat calls for writing. Calibrated decisions
Learning extracts lessons after /done or at exit, in your own words, into files under learning/ / /; later sessions recall the ones that match. With [supervisor.learning.evolution] enabled, recurring lessons become skill and guardrail candidates that run in shadow first, then an A/B trial against a control arm, and go active only when they beat it. Learning