RightNow-AI/openfang

An open-source "agent operating system" written in Rust that compiles to a single binary and runs autonomous agents on schedules.

What it is

OpenFang is a self-hosted runtime for autonomous LLM agents, built in Rust and shipped as a single binary. Rather than only answering when prompted, it runs pre-packaged autonomous capabilities called "Hands" on schedules — for research, lead generation, target monitoring, social media management, and similar recurring work — and reports to a local dashboard. It targets developers who want to run long-running, self-directed agents locally instead of assembling a Python agent stack.

Key features

Hands: autonomous capability packages that run independently on schedules. Seven are bundled (Clip, Lead, Collector, Predictor, Researcher, Twitter, Browser), each defined by a HAND.toml manifest, a multi-phase system prompt, a SKILL.md, and approval guardrails.
40 channel adapters (Telegram, Discord, Slack, WhatsApp, Signal, Matrix, email, and more) with per-channel model overrides, DM/group policies, rate limiting, and output formatting.
27 LLM providers reached through 3 native drivers (Anthropic, Gemini, OpenAI-compatible), with routing by task-complexity scoring, automatic fallback, and cost tracking.
OpenAI-compatible API: 140+ REST/WS/SSE endpoints and a built-in web dashboard at http://localhost:4200.
Security systems (16 listed), including a WASM sandbox with fuel metering, a Merkle hash-chain audit trail, Ed25519-signed agent manifests, SSRF protection, and a GCRA rate limiter.
Memory and migration: SQLite persistence with vector embeddings, plus a migration engine that imports agents, history, skills, and config from OpenClaw, LangChain, and AutoGPT.

Tech stack

Rust (edition 2021, rust-version = 1.75), organized as a Cargo workspace of 14 crates.
Async runtime tokio 1; HTTP server axum 0.8 with tower/tower-http; HTTP client reqwest 0.12 over rustls.
WASM sandbox via wasmtime 43.
Storage via rusqlite 0.31 (bundled SQLite).
Security stack: ed25519-dalek 2, aes-gcm 0.10, argon2 0.5, hmac, sha2, zeroize, and governor 0.10 for rate limiting.
MCP via rmcp 1.2 (the official Rust SDK); CLI/TUI via clap 4 and ratatui 0.29.
Email through lettre 0.11 and imap; MQTT through rumqttc.
Desktop app built on Tauri 2.0; optional WhatsApp Web gateway needs Node.js >= 18.
Workspace Cargo.toml declares license Apache-2.0 OR MIT.

When to reach for it

You want agents that run on a schedule and act on their own (monitoring, research, lead generation), not just interactive chat.
You prefer a single self-hosted binary over managing a Python environment and its dependencies.
You need many messaging-channel integrations available out of the box.
You want a local OpenAI-compatible endpoint backed by agent routing and cost tracking.

When not to reach for it

You need a stable, pinnable API: the project is pre-1.0 and its own notice warns of breaking changes between minor versions, recommending you pin to a specific commit for production.
Your team extends agents in Python and expects that ecosystem; the core is Rust.
You only need a simple prompt/response chatbot — the operating-system model is heavier than that use case requires.
You depend on features from the less-mature Hands; the README flags Browser and Researcher as the most tested, implying others are less so.

Maturity signal

Created in February 2026, OpenFang has drawn roughly 18,000 stars in a few months and receives frequent pushes (last push July 2026), so it is actively developed with strong early traction. But it describes itself as feature-complete yet pre-1.0 (Cargo.toml version 0.6.9), and its own stability notice warns of breaking changes and advises pinning to a commit for production — so treat it as fast-moving rather than API-stable. Licensing is stated inconsistently across sources: repository metadata lists Apache-2.0, the workspace manifest declares Apache-2.0 OR MIT, and the README's license section says MIT.

Alternatives

LangGraph — use instead when you are already in the Python/LangChain ecosystem and want graph-structured orchestration.
CrewAI — use instead when you want a Python multi-agent framework organized around role-based crews.
AutoGen — use instead when you want a Python conversational multi-agent framework and Docker-based isolation.

Notes

The entire system, including the bundled Hands and skills, compiles into one binary the README cites at about 32 MB — no pip install or Docker pull to add capabilities.
The comparison tables and benchmark charts in the README ("Measured, Not Marketed") are self-reported; the numbers warrant independent verification.
Version numbers are inconsistent within the README itself (a v0.5.10 notice alongside a 0.6.9 badge and a 0.6.9 workspace version).

Wiki: RightNow-AI/openfang