Octomind CLI
Octomind is a coding agent in one Rust binary: safety rules in TOML instead of approval prompts, with a supervisor model on top
- 657 ★ on the repository
- Last commit
- Apache-2.0
Broad access
Why this level
- The agent reads and edits files and runs shell commands in your repository, and the guards depend on the rules you write
- In daemons and CI the agent works without a person until rules and budget stop it
- Provider keys sit in its own `.env` file, and the supervisor sends part of the context to a second model
Install
Manual install
curl -fsSL https://raw.githubusercontent.com/muvon/octomind/master/install.sh | bashFor macOS and Linux. Read the script before running it.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent reads and edits files and runs shell commands in your repository, and the guards depend on the rules you write
- In daemons and CI the agent works without a person until rules and budget stop it
- Provider keys sit in its own `.env` file, and the supervisor sends part of the context to a second model
This is third-party code. Review the repository files before installing.
What it does
Octomind is an open coding agent harness in a single Rust binary. It gives any model the tools to work in a repository: reading and editing files, running a shell, code search and delegation to subagents through MCP. The same agent runs interactively, over stdin, as a daemon, over WebSocket and over ACP, and its settings live in one TOML file. Safety rules live in `.agents/guardrails.toml`. Pre-call guards deny commands by pattern, for example `rm -rf`, and conditional rules allow `git push` only after `git status`. Hooks after a result feed linter failures back to the agent, and validators check a turn after it ends. This replaces "approve this change?" questions where the agent runs without a person. A second supervisor model keeps the plan, reminds the goal and trims oversized tool results. `octomind run developer` loads a ready specialist from a tap registry with a prompt, model, MCP servers and permissions. Skills come from standard `SKILL.md` folders, and after `/done` the agent writes lessons in your own words and recalls them in later sessions.
Who it is for. Developers who run coding agents in CI, on schedules and as daemons and want to set limits in code instead of clicking approve.
Good fit when
- An agent runs in CI without a person and dangerous commands such as `rm -rf` must be denied by pattern
- An autonomous loop should run the linter after each result and feed errors back to the agent
- A long task runs into context limits and old turns should be compressed without losing the ability to recall them
- A session should live as a daemon that you send commands to from another terminal with `octomind send`
- A model should be chosen per role with spending capped per request and per session
Not a fit when
- You need an add-on for Claude Code or Cursor: Octomind is a separate harness with its own interface
- You want a rich pane-based interface: the session runs in a plain line prompt with Markdown rendering
- TOML rules are not enough: complex checks have to be written as hook and validator scripts
Example request
Run octomind run developer in this repository, fix the two failing tests and add a rule to guardrails.toml that blocks git push without git statusLimitations
Octomind replaces the agent rather than extending Claude Code. On first run you must choose a model source: the Octomind hub for credits you buy, your own provider key (OpenRouter, OpenAI, Anthropic, DeepSeek, Ollama) or the shipped defaults. The key is saved in its own `.env` readable only by you. The supervisor intent check is off by default and spends extra tokens. Installation runs a `curl ... | bash` script, or `irm ... | iex` on Windows, and you should read the script before running it. The comparison with other agents in the README was made by the authors.
How to disable. Delete the `octomind` binary and the configuration with `.env`. A Homebrew install is removed with `brew uninstall octomind`. A daemon is stopped by ending the session under the name given at launch.
FAQ
How do I install Octomind?
On macOS and Linux run `curl -fsSL https://raw.githubusercontent.com/muvon/octomind/master/install.sh | bash`, on macOS you can use `brew install muvon/tap/octomind`, and Windows PowerShell has `install.ps1`. Then start `octomind`.
How do I start a ready agent?
Run `octomind run developer` in the project folder. The specialist and its settings come from the `muvon/tap` registry, not from your local config.
How do I set safety rules?
Put `.agents/guardrails.toml` in the repository: `[[guard]]` blocks deny calls before they run by pattern, with an optional `when` condition. Hooks and validators add checks after a result and after a turn.
Which models can I use?
The Octomind hub or your own key: OpenRouter, Anthropic, OpenAI, DeepSeek, Ollama and others. A model is picked per role, and spending is capped per request and per session.
Can I run Octomind without a UI?
Yes: `echo "..." | octomind run developer --format plain` for scripts and CI, while `--daemon` keeps a session alive and `octomind send` talks to it.
Discussion
No questions or reviews yet.
Related
CLI toolsA terminal coding agent in one Rust binary that reads code, edits files and runs tests with any model of your choice.
A Rust CLI where a meta-agent hands work to Claude Code and Codex, each on its own branch and tmux session, in one view.
A Rust terminal coding agent with a web interface, subagents, memory, skills and MCP, running on GPT and Claude models.
A Rust host for coding agents: a dashboard, required checks, review by another model and a signed receipt for each turn
Локальная среда для команды агентов: проекты как каталоги, промпты как файлы и журнал каждого шага, который можно проверить.
A lightweight Rust coding agent for the terminal, inspired by pi and OpenCode and built for a small memory footprint
README in short
The README describes Octomind as an open coding agent harness in one binary and shows the turn flow: what is assembled before a request, the tool loop, compression, guards and the supervisor. It then gives reasons to choose it, a quick start, benchmark results with a reproducible method and a section on safety rules as code. The end collects launch surfaces, installation, configuration and documentation links.