OpenFang

Overview

Summary + Metrics + Quick Start

Overview
What It Is
Core Concepts
Architecture
Security (16 Layers)
Hands
Ecosystem
Getting Started
Benchmarks
Whitepaper
Sources
OpenFang · Agent Operating System
This is not 'just another Agent framework', but a system treating long-running Agents as first-class citizens: lifecycle hosting, layered security, observability, and reusable Skills/Tools ecosystem.
Quick StartView Blog →Skills Lib →
OpenFang vs traditional frameworks

Overview

OpenFang is an open-source Agent Operating System built entirely from scratch in Rust. Its goal is not to be just another "conversational agent framework," but rather to provide a runtime and governance layer closer to a "system/kernel":

  • Long-Running: Agents/Hands work persistently on a schedule, rather than waiting for you to send a message.
  • Strong Isolation & Controlled Execution: Tool code runs preferentially in a WASM sandbox and features dual-metering (fuel + epoch).
  • Auditable: Links critical actions into a verifiable audit path via a Merkle hash-chain.
  • Governance: Capability-based permissions, RBAC, quotas and billing, rate limiting, and policy-driven channel access.
  • Unified Delivery: Multi-crate stratification that ultimately compiles down to a single binary (~32MB according to README/official examples).

Key Metrics

MetricValue (Public Data)
Codebase Size~137K lines of Rust
Workspace Crates14 crates (13 code + xtask)
Tests1,700+
DistributionSingle binary (~32MB according to official docs)
Channels40+ (docs/architecture states 40)
LLM Providers20+ (Official/README show 26~27)
Security Layers16 layers (docs/security)

One-Sentence Difference: Why Is It Called an "Agent OS"?

Traditional agent frameworks act more like "libraries": You write code, call a flow, send a message, and wait for a result.

OpenFang acts more like a "system":

  • It has a "Kernel" (openfang-kernel) responsible for assembling subsystems, scheduling, RBAC, metering, workflows/triggers, and graceful shutdowns.
  • It has a "Runtime" (openfang-runtime) handling the agent loop, LLM drivers, tool runners, WASM sandbox, and MCP/A2A.
  • It has a "Persistent Substrate" (openfang-memory) built on SQLite for sessions/memory/embedding/KG/canonical sessions.
  • It has an "Integration Layer" (openfang-channels) connecting agents to Telegram/Slack/Discord/... applying per-channel policies, formatting, and rate limiting.

Quick Start (Shortest Path)

  1. Install:
bash
curl -fsSL https://openfang.sh/install | sh
  1. Initialize:
bash
openfang init
  1. Configure a provider key (example):
bash
export ANTHROPIC_API_KEY=sk-ant-...# or export OPENAI_API_KEY=...
  1. Start daemon + Open dashboard:
bash
openfang start# Dashboard: http://127.0.0.1:4200

The commands above are taken from the official Getting Started documentation.