Octomind.run

Octomind is an open-source coding agent harness and agent runtime built by Muvon. It gives any LLM tools to read and edit files, run shells, search code and 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 an editor. It is released under Apache 2.0, with latest release 0.70.1.

What It Is

Octomind is a single Rust binary that wraps a model in a session engine with tools, roles, guardrails and a supervising model. An agent is defined in TOML (model, tools, prompts, guardrails), so the same agent behaves the same on a laptop, a cloud machine or a CI pipeline. Running a specialist such as octomind run developer:general fetches its manifest from a tap registry, installs required tools, and opens a session.

How a Session Works

Context is assembled per user message from AGENTS.md, matching skills (standard SKILL.md folders) and recalled lessons from earlier sessions. A second supervisor model keeps the plan, trims oversized tool results and refuses a self-reported "done" until evidence checks out. Adaptive compression folds older turns into summaries while raw messages stay in a per-session archive that a recall tool can restore verbatim. Sessions can be named and resumed across days.

Policy as Code

Instead of approval popups, a .agents/guardrails.toml file defines pre-call guards, post-result hooks and post-turn validators that run deterministically, so agents can run unattended in CI, daemons or scheduled jobs. An optional --sandbox flag restricts writes to the working directory using Landlock on Linux or Seatbelt on macOS. Spending caps can be set per request and per session; non-interactive runs stop with exit code 75 when over budget.

Models, Surfaces and Extensions

Models can come from the Octomind hub or from your own keys (OpenRouter, Anthropic, OpenAI, Google, DeepSeek, Bedrock, Ollama and others) and can be swapped mid-session with /model. Extension is via MCP servers, local scripts in .agents/tools/, skills and plugins; the tap system shares installable agents, workflows and skills. Structured output is available as JSONL or JSON Schema.

Benchmark Claims

The project reports that on 25 tasks from merged pull requests in 2026, Octomind with glm-5.2 solved 24, versus 19 for OpenCode with the same model, and 23 for Claude Code with claude-opus-5. The authors publish a reproduction guide (octobench).