Octomind: Open-source Rust AI agent harness outperforms Claude Code on benchmarks

• Octomind is an open-source, Rust-based AI coding agent runtime developed by Muvon, released under the Apache 2.0 license with 2,733 commits.
• In the octobench benchmark (July 30, 2026), Octomind paired with glm-5.2 solved 24 of 25 tasks with a judge score of 2264/2500, costing $63.43 and taking 3.6 hours.
• This performance exceeded Claude Code with claude-opus-5, which solved 23 of 25 tasks with a score of 2262/2500 at a cost of $81.79 over 6.7 hours.
• The tool features a deterministic guardrail system using TOML-based guards, hooks, and validators to enforce safety without interactive approval popups.

Architecture and Features

• Built as a single binary with native MCP support, allowing LLMs to read/write files, execute shell commands, and delegate to sub-agents.
• Supports five execution surfaces: interactive CLI, piped stdin, background daemon, WebSocket server, and ACP sub-agent mode.
• Implements adaptive compression to reduce token usage while maintaining task knowledge and real-time context.

Configuration and Deployment

• Uses a "Policy as Code" approach where all settings, including models, tools, and budgets, are managed via TOML files.
• Supports Bring Your Own Key (BYOK) for providers like OpenRouter, Anthropic, OpenAI, and DeepSeek, or Octomind Cloud.
• Enforces spending limits per request or session, terminating non-interactive runs with exit code 75 if thresholds are exceeded.