rufous

rufous is an alpha-release agent harness written in Rust, designed around the limits of small local models (2B–12B parameters). It keeps a Record of every Tool Call the model makes, which the model cannot edit, and before a Turn ends it compares the model's report against that Record, holding the reply at most once per Turn if files named by the user were not written or a declined write was attempted. Every Tool Call passes through a five-ring Safety Gate that asks the user when present and refuses when not. It keeps a small system prompt (about 1,250 tokens) with nine core tools and enables additional tools progressively. It runs as a single binary for macOS, Linux and Windows, offers a terminal interface, a local web UI (rufous serve), and an editor agent over the Agent Client Protocol (rufous acp, tested with Zed). It connects to Ollama, LM Studio, llama-server, or any OpenAI-compatible endpoint, with optional cloud providers. It is licensed under Apache-2.0 and the page states no account is required. The page reports early, small-sample measurements on qwen3.5:4b and gemma4:e4b, and notes a desktop app is still in development and not in the alpha.

Key features

• Record of every Tool Call that the model cannot edit
• Checks the model's reply against the Record before a Turn ends
• One Hold per Turn at most; honest 'I couldn't' replies always pass
• Five-ring Safety Gate for Tool Calls, including permission prompts and refusals
• Progressive tool disclosure via enable_tools, with skills loaded on use
• Per-folder instructions via SYSTEM.md or AGENTS.md
• Web UI via rufous serve with sessions, permission cards and live plan
• Terminal chat and unattended task-file runs via rufous run
• Editor agent over Agent Client Protocol (rufous acp), tested with Zed
• rufous doctor command for checking setup
• Offline mode that disables cloud provider, web tools and remote MCP servers
• Connects to Ollama, LM Studio, llama-server and OpenAI-compatible endpoints
• Single binary install script with SHA-256 checksum verification

Best suited for

• Running small local language models (2B–12B) on your own machine
• Developers who want an agent in their editor (such as Zed) backed by a local model
• Users who want agent work that checks file writes against a tamper-resistant record
• Users who need an offline or privacy-focused local agent setup

Pricing

Not stated on the website