What it does

OpenFang is a single ~32MB Rust binary that runs autonomous agents 24/7. Instead of waiting for prompts, its “Hands” wake up on schedule to research competitors, generate leads, monitor targets for OSINT-style intelligence, manage social media, or even edit and publish video clips. Everything compiles in; no pip install, no Docker pull, no downloading models at runtime.

The interesting bit

The “Hand” packaging is the clever part: each bundles a HAND.toml manifest, a 500+ word multi-phase system prompt, a SKILL.md domain reference, and guardrails with approval gates for sensitive actions like purchases. It’s basically an operating system process model, but for LLM agents—complete with pause/resume state and a dashboard at :4200.

Key highlights

• Seven built-in Hands: Clip (video editing pipeline), Lead (ICP scoring), Collector (OSINT monitoring), Predictor (superforecasting with Brier score tracking), Researcher (APA-cited reports), Twitter (content + approval queue), Browser (Playwright automation with mandatory purchase gates)
• 137K LOC across 14 crates, 1,767+ tests, zero Clippy warnings
• 40 channel adapters, 27 LLM providers, 16 “security systems” (per their own comparison matrix)
• Self-reported benchmarks: 180 ms cold start, 40 MB idle memory—faster and lighter than Python frameworks, though ZeroClaw edges it on raw speed

Caveats

• Pre-1.0 with explicit warning: “expect rough edges and breaking changes between minor versions”
• Self-published benchmarks against competitors; methodology and exact definitions of “security systems” aren’t independently verified
• Feature comparison table in README is truncated, so full capability matrix is unclear

Verdict

Worth a look if you want agents that actually run on cron schedules rather than chat loops, and you prefer a compiled binary to a Python stack. Skip it if you need stability guarantees today or want to deeply customize agent logic beyond HAND.toml configuration.