OPEN SOURCEMIT LICENSEBUILT IN RUST

OpenFang

the agent operating system.

open-source Agent OS built in Rust. 14 crates, 137K lines of code, zero clippy warnings. deploy autonomous AI agents called Hands that run on schedules, build knowledge graphs, and report to your dashboard. battle-tested architecture.

install

curl -fsSL https://openfang.sh/install | sh

[visit openfang.sh][github repo][@openfangg]

stats.overview()

autonomous hands

38+

built-in tools

channel adapters

LLM providers

security systems

140+

REST API endpoints

hands.explain()

Hands are autonomous capability packages. each Hand is a specialized AI agent with its own identity (SOUL.md), memory, schedule, and tools. they run continuously, persist state across sessions, and can be chained together. think of them as departments in an autonomous company.

collector

OSINT-style intelligence monitoring. change detection, sentiment analysis, knowledge graph construction.

researcher

fact-checking engine. CRAAP methodology, cross-referenced sources, cited reports with confidence scores.

twitter

autonomous X/Twitter operator. scheduled posting, engagement automation, narrative tracking.

lead

autonomous lead generation with ICP scoring, knowledge graphs, and qualification pipelines.

clip

converts long-form video into short-form content with auto-captions and thumbnail generation.

predictor

forecasting engine using Brier score calibration. probability estimates with tracked accuracy.

browser

web automation. form-filling, multi-step workflows, data extraction from any website.

architecture.deep_dive()

runtime

WASM sandboxed execution

tool code runs inside WASM with dual metering (fuel + epoch interruption). file operations are workspace-confined. each Hand gets its own isolated execution environment with resource limits.

[runtime]
sandbox    = "wasm"
metering   = "fuel + epoch"
file_ops   = "workspace-confined"
isolation  = "per-hand"

memory

persistent knowledge system

SQLite-backed storage with vector embeddings. cross-channel canonical sessions, automatic LLM-based compaction, and JSONL session mirroring. agents remember everything across restarts.

[memory]
backend    = "sqlite + vectors"
sessions   = "cross-channel"
compaction = "llm-based"
mirroring  = "jsonl"

workspace

Hand file structure

every Hand gets a standardized workspace with identity, memory, and session files. SOUL.md defines the agent's core purpose. IDENTITY.md holds personality metadata. memory persists as dated markdown logs.

workspace/
├── SOUL.md        # core identity
├── IDENTITY.md    # personality
├── BOOTSTRAP.md   # first-run protocol
├── MEMORY.md      # persistent knowledge
├── memory/        # dated logs
└── sessions/      # conversation history

desktop

Tauri 2.0 native app

native desktop application built with Tauri 2.0. system tray integration, native notifications, single-instance enforcement, and auto-start on boot. runs on macOS, Linux, and Windows.

[desktop]
framework  = "tauri 2.0"
tray       = true
autostart  = true
platforms  = ["macos", "linux", "windows"]

security.audit()

16 security systems built-in. not bolted on — designed from the ground up. every layer is there because agents with tools need guardrails.

●WASM dual-metered sandbox (fuel + epoch)

●Ed25519 manifest signing

●Merkle audit trail

●taint tracking across data flows

●SSRF protection

●secret zeroization on drop

●HMAC-SHA256 mutual auth

●GCRA rate limiter

●subprocess isolation

●prompt injection scanning

●workspace-confined file ops

●encrypted vault storage

channels.list()

40 channel adapters. your agents communicate wherever your users are.

TelegramDiscordSlackWhatsAppTeamsIRCMatrix+ 33 more adapters

protocols.supported()

interoperability-first. agents talk to tools, other agents, and the network.

MCP

Model Context Protocol — tool integration standard

A2A

Google Agent-to-Agent — inter-agent task delegation

OpenFang Protocol

P2P networking between OpenFang instances

benchmarks.compare()

compared against CrewAI, AutoGen, LangGraph, and others.

metric	openfang	others
cold start	180ms	up to 5,980ms
security depth	16 systems	1–6 systems
channel adapters	40	0–15
LLM providers	27 native	3–12

fanghand runs on OpenFang

the fanghand marketplace is built entirely on OpenFang v0.3+.
every Hand you see in the marketplace is an OpenFang agent.
collector, researcher, twitter — all running as autonomous Hands.

[browse the marketplace][build your own Hand →]