Openfang
Open-source Agent Operating System
Last verified:
What is Openfang?
OpenFang is an open-source Agent Operating System built entirely in Rust, designed to deploy and manage autonomous AI agents that work for you on schedules rather than waiting for prompts. Unlike traditional agent frameworks that function as chatbot wrappers, OpenFang is a full operating system with 137K lines of Rust code, compiling to a single 32MB binary with zero runtime dependencies. It features 7 autonomous
Openfang pricing
Pricing model: Freemium
OpenFang is completely free and open-source under the MIT license. The software itself has no cost — you can install, use, and modify it without paying. The only costs are LLM provider API keys (Anthropic, OpenAI, Gemini, Groq, DeepSeek, etc.) which you provide yourself, or you can run local models via Ollama for free. Optional VPS hosting costs approximately $5/month at Hetzner if you want to run it server-side. Users can begin with preloaded AI credits or utilize their own API keys for comprehensive control over models and expenses.
Openfang pros
- Open-source with MIT license, free to use
- Single 32MB binary with zero runtime dependencies
- 16 security layers including WASM sandbox and Merkle audit trail
- 7 autonomous Hands that work on schedules without prompting
- 40 built-in channel adapters for all major messaging platforms
- 27 LLM providers with 123+ models and intelligent routing
- Cold start time of 180ms vs 3-6 seconds for Python frameworks
- 40MB idle memory usage vs 180-400MB for competitors
- 53 built-in tools plus MCP and A2A protocol support
- SQLite-backed persistent memory with vector embeddings
- Tauri 2.0 native desktop app with system tray and notifications
- 1,767+ passing tests with zero clippy warnings
- 30 pre-built agent templates across 4 performance tiers
- 60 bundled skills with FangHub marketplace
- OpenAI-compatible API for drop-in replacement
- Hot-reloadable configuration without restart
- Built-in migration tool from OpenClaw
- 140+ REST/WebSocket API endpoints including SSE
Openfang cons
- Requires LLM provider API key (costs not included)
- New project with smaller community than established frameworks
- Rust-based, steeper learning curve for Python developers
- Only supports x86_64 and ARM64 architectures
- Minimum 50MB disk space plus workspace storage
- Requires macOS 12+, Linux glibc 2.17+, or Windows 10+
- Autonomous Hands may require careful configuration to avoid unwanted actions
- Purchase approval gate required for Browser Hand transactions
Frequently asked questions about Openfang
How does OpenFang ensure AI agent security?
OpenFang combines 16 discrete security systems including WASM dual-metered sandbox (fuel + epoch interruption), taint tracking for data leaks, Ed25519-signed agent manifests, Merkle hash-chain audit trail for tamper-evident logging, HMAC-SHA256 mutual authentication for P2P, SSRF protection, secret zeroization, capability-based access control with inheritance validation, subprocess isolation, prompt injection scanner, and path traversal prevention.
What distinguishes Hands from regular AI agents?
Hands are OpenFang's core innovation — pre-built autonomous capability packages that run independently on schedules without user prompts. Traditional agents wait for you to type, but Hands work for you autonomously: the Researcher fact-checks with CRAAP evaluation, Lead generates qualified leads daily, Collector monitors targets continuously, Predictor makes forecasts with Brier score tracking, Twitter manages your X account, Browser automates web workflows, and Clip turns video into shorts. Each Hand bundles a HAND.toml manifest, multi-phase system prompt, SKILL.md expert knowledge, and dashboard metrics.
Can OpenFang integrate with existing LLMs like GPT-4 or Claude?
Yes, OpenFang supports 27 providers with 123+ models including Anthropic (Claude), OpenAI (GPT-4), Gemini, Groq, DeepSeek, and more. It features intelligent routing, cost tracking, per-channel model overrides, and cost controls. You can set your API key via environment variables (ANTHROPIC_API_KEY, OPENAI_API_KEY, GROQ_API_KEY, etc.) during initialization, or use local models via Ollama.
Is OpenFang suitable for Windows environments?
Yes, OpenFang compiles to a single 32MB binary for macOS 12+, Linux (glibc 2.17+), and Windows 10+. Installation is via one-liner: curl -fsSL https://openfang.sh/install | sh on Linux/macOS or irm https://openfang.sh/install.ps1 | iex on Windows. The binary lands in ~/.openfang/bin and gets added to PATH. Direct downloads are available for x86_64 and ARM64 architectures in .tar.gz, .zip, .dmg, and .msi formats openfang +2 .
How does the Merkle audit trail enhance compliance?
The Merkle audit trail hashes all agent actions into an immutable chain, enabling cryptographic verification of activity logs for regulatory audits or incident analysis. Combined with Ed25519 signed agent manifests and a 10-phase graceful shutdown, it creates tamper-evident logging that prevents code injection and ensures all agent actions are trackable and auditable.
What channels does OpenFang support?
OpenFang has 40 built-in channel adapters covering Telegram, Discord, Slack, WhatsApp, Teams, IRC, Matrix, Signal, and 34 more platforms. Each channel supports per-channel model overrides, DM/group policies, rate limiting, and output formatting. This means one agent can operate across all major messaging platforms simultaneously.
How do I install and get started with OpenFang?
Install with one command: curl -fsSL https://openfang.sh/install | sh. Then run openfang init to walk through provider setup and create your first agent. Set your LLM API key (e.g., export ANTHROPIC_API_KEY=sk-ant-...). Start the daemon with openfang start — the dashboard becomes live at http://localhost:4200. Activate a Hand with openfang hand activate researcher or spawn a custom agent with openfang agent spawn coder. The entire setup takes under 2 minutes openfang +2 .
What performance benchmarks does OpenFang achieve?
OpenFang achieves 180ms cold start (vs 2500-5980ms for others), 40MB idle memory (vs 180-394MB), and 32MB install size (vs 100-500MB). It has 16 security systems (vs 1-6 for competitors), 40 built-in channel adapters (vs 0-15), and 27 native LLM providers (vs 8-28). The codebase has 1,767+ passing tests with zero clippy warnings and 14 Rust crates with modular kernel design.
Can I build my own custom Hands?
Yes, you can define a custom HAND.toml with tools, settings, requirements, and a system prompt to build your own Hand. The Hand development guide is available in the documentation. Once built, you can publish your custom Hand to FangHub marketplace for others to use. Use openfang hand --help to see all commands including activate, deactivate, status, list, pause, and resume.