Octomind – open-source AI harness that saves your tokens, static rust

Who Should Read

Backend or DevOps developers who want to run coding agents in CI or daemons without human approval prompts.

Core Mechanics

• One binary, any model
  • Octomind is an open-source coding agent harness that lets an LLM read and write files and run shells in your repo.
  • A harness is the runtime wrapper that manages tools and sessions around a model.
  • It works with OpenRouter, Anthropic, OpenAI, DeepSeek, Ollama and more, with per-role model choice.
• Five ways to run it
  • Interactive CLI, stdin pipe, daemon (commands via octomind send), WebSocket and ACP (editor protocol) are supported.
  • They share the same engine and config.
• TOML guardrails instead of approval popups
  • Put regex deny rules (guards) in .agents/guardrails.toml.
  • Hooks run after tool results and validators run after each turn.
  • The README maps these to Claude Code's PreToolUse, PostToolUse and Stop hooks.
• Own benchmark: 24 of 25 solved
  • Tested on 25 tasks from merged PRs in Python, PHP, Rust, C++ and JavaScript.
  • octomind+glm-5.2 solved 24/25 at $63.43 in 3.6h.
  • claude code+claude-opus-5 solved 23/25 at $81.79 in 6.7h.
• Harness mattered with the same model
  • opencode with the same model and endpoint solved 19/25 at $129.54.
  • The README says octomind solved 24 vs 19 at roughly half the cost.
  • codex+gpt-5.6-sol solved 21/25 at $14.86 in 1.0h, the cheapest and fastest.
• Budget overruns have defined behavior
  • You can cap spend per request and per session.
  • A terminal session asks y/N when over budget.
  • A non-interactive run stops with exit code 75.

Community Reactions

• Benchmarks called outdated
  • One commenter said the comparison used older models than Opus 5.5 and GPT 6.1 Sol.
  • They noted thinking effort and Claude Code, Codex and OpenCode versions are not specified.
• Does quality suffer?
  • One commenter said tokens are cheap and quality is what matters, asking whether quality goes up or down.
• Differentiation unclear
  • One commenter felt the README reads like AI-written text and saw no clear difference from other harnesses.
  • Others asked for comparisons with maki.sh and oh-my-pi.
• Concern about curl installers
  • One commenter asked for installation via Linux package managers because curl scripts feel insecure.

How to Apply

• Block risky commands with a guardrail file
  • Create .agents/guardrails.toml in a test repo with an rm -rf deny rule.
  • Run octomind run developer:general and ask the agent to delete files to confirm it is blocked.
• Run the agent in CI via stdin pipe
  • Run one review or explain task with echo "..." | octomind run developer:general --format plain.
  • Handle exit code 75 for budget overruns in the CI step.
• Benchmark agents on your own repo
  • Take 3-5 past PRs from your repo and compare your current harness with octomind on the same model.
  • Record solves, cost and wall time.

Try It

# Run in a shell (macOS/Linux). If curl scripts worry you, use brew or cargo install octomind.

brew install muvon/tap/octomind
export OPENROUTER_API_KEY="your_key"

# Write guardrails at your project root
mkdir -p .agents
cat > .agents/guardrails.toml <<'EOF'
[[guard]]
match = "shell(command=^rm\\s+-rf?)"
message = "rm -rf blocked."

[[guard]]
match = "shell(command=git push)"
when = ["+shell(command=git status)"]
message = "Review changes before pushing."
EOF

# Non-interactive run
echo "Explain the auth module" | octomind run developer:general --format plain

# Next: in a session, ask it to run rm -rf and to git push without git status, and watch the blocks.

Terminology

Related Papers

• Postman silently auto-installs AI plugins and attempts to share all API data

  Opening Postman auto-installed plugins and a startup hook into Codex and Claude Code

• I Am the Reason OpenAI Is Killing Unlimited Chats

  Connecting local MCP tools to regular ChatGPT chats gave a parallel coding-agent pool without Codex limits

• Toward Bounding Coding-Agent Autonomy in Software Maintenance: A Pilot Study of Boundary Contracts

  Telling a coding agent each file's role upfront nearly doubled strictly acceptable patches, from 9 to 18 of 40.

• GPT-5.6 Sol can beat Factorio on peaceful mode

  GPT-5.6 Sol in Codex at low thinking finished a peaceful Factorio run with a rocket launch

• Do the math: How we took over ChatGPT accounts

  One public share link ran attacker JS in a victim's ChatGPT session and could leak private chats

• AI isn't replacing SREs, it's deskilling them

  Letting AI handle 95% of incidents may dull the skills SREs need for the critical 5%.

• Backdoor in Claude Code for VS Code extension on the Cursor Marketplace

  A tampered Claude Code extension on Cursor Marketplace planted an SSH backdoor

• Building an AI-native engineering org

  Automating each shipping step raised merges per engineer about 5x with reverts under 0.5%

Related Resources

• https://github.com/Muvon/octomind
• https://github.com/Muvon/octomind#benchmarks
• https://openrouter.ai/rankings
• https://github.com/highflame-ai/codeoid