Muvon/octomind

What it does

Octomind is a coding-agent harness written in Rust. It gives an LLM repository tools — reading and editing files, running shells, searching code, delegating to sub-agents — over MCP, and exposes the same session through an interactive CLI, stdin pipe, non-interactive --format output, a daemon, a WebSocket server, or an ACP agent inside editors. The core design choice is policy as code: deterministic guards, hooks, and validators live in .agents/guardrails.toml and scripts rather than approval prompts, so sessions can run unattended in CI and scheduled jobs. A second supervising model keeps the plan, trims oversized tool output, and refuses a self-reported "done" until evidence checks out. Optional learning extracts lessons after /done and can promote recurring lessons into skill or guardrail candidates. For a deeper tour, the guardrails guide and supervisor docs explain the two most distinctive pieces.

Quick start

Install on macOS or Linux with the install script, or on Windows with the PowerShell script. Homebrew is also available on macOS:

curl -fsSL https://raw.githubusercontent.com/muvon/octomind/master/install.sh | bash
# or: brew install muvon/tap/octomind

Then get model access. The shortest path is Octomind Cloud, which stores a gateway key locally:

octomind login

To use your own provider key instead, export one (for example OPENROUTER_API_KEY, OPENAI_API_KEY, ANTHROPIC_API_KEY, or DEEPSEEK_API_KEY) and point the model tables at it; the README notes that supervisor and compression profiles also default to the cloud gateway, so update all three. Now run a specialist from your project root:

cd ~/your/project
octomind run developer:general

You land in a plain line prompt where the agent can read code, run commands, and edit files. The first run of a specialist may install its tools and ask for credentials. For a non-interactive run, pipe a prompt instead: echo "Explain the auth module" | octomind run developer:general --format plain. A recorded bug-fix example walks through failing tests, an agent fix, and an independent test rerun.

Requirements and limits

• The install script targets Linux and macOS on x86_64/aarch64 and installs to ~/.local/bin/; the PowerShell script installs to %LOCALAPPDATA%\octomind\bin and modifies the user PATH. Cargo builds need Rust 1.95+ and protoc.
• A model provider is mandatory: either Octomind Cloud via octomind login or a provider API key. Every role has its own model and temperature, so cost and behaviour depend on the configuration you write.
• Spending can be capped per request and per session. Over budget, an interactive session asks y/N; a non-interactive run exits with code 75.
• Sandboxing is opt-in: --sandbox locks writes to the working directory via Landlock on Linux or Seatbelt on macOS, inherited by shells, MCP servers, and sub-agents. Without it, tool writes are governed by your guardrails and operating-system permissions.
• Anonymous telemetry is on by default and excludes code, prompts, paths, and tool arguments; opt out with DO_NOT_TRACK=1, OCTOMIND_TELEMETRY=0, or telemetry = false in config.
• The project is Apache-2.0 licensed.

Benchmark claims

The README reports a July 30, 2026 run on octobench: 25 tasks harvested from merged pull requests, each solved in the pre-fix repository and judged by held-out tests. Octomind with glm-5.2 solved 24/25 at $63.43 and 3.6h; Claude Code solved 23/25, Codex 21/25, and OpenCode with the same glm-5.2 model 19/25 at $129.54. The per-case table and reproduction guide are in BENCHMARK.md. These are published results from an older model set, not a measurement of the current checkout.

HotRepo's take

The repository appeared on HotRepo's Rust daily list on 2026-11-04 with 126 stars gained and a best position of 13, and an HN thread the day before drew 54 points and 37 comments. That discussion raises a concrete concern: one commenter notes the published benchmark compares Opus 5, GPT-5.6 Sol, and GLM-5.2 and does not list thinking-effort settings or the versions of the competing agents (discussion). The maintainer replied that updated frontier-model benchmarks are in progress. So treat the benchmark numbers as a July snapshot and watch for a refreshed table before drawing conclusions about current relative performance. The practical differentiators to evaluate are the TOML guardrail model and the five-surface deployment story, both of which are documented in depth rather than asserted only in the README.