OpenFang

OpenFang is an open-source Agent Operating System maintained by the RightNow-AI organization, built in Rust (137,728 lines across 14 crates) and delivered as a single ~32MB binary. Current version v0.5.10, pre-1.0 but marked feature complete.

Core Innovation: Hands Mechanism#

Pre-built autonomous capability packages that run on schedule without manual prompting. Each Hand contains a HAND.toml manifest, 500+ word multi-stage system prompts, SKILL.md domain knowledge injection, and safety guardrails, all compiled into the binary:

• Clip: YouTube URL → download → best segment identification → vertical short-video clipping (subtitles/thumbnails/AI narration) → publish to Telegram/WhatsApp, 8-stage pipeline
• Lead: Daily automatic ICP-matched prospect discovery, web research enrichment, 0-100 scoring, deduplication, CSV/JSON/Markdown output
• Collector: OSINT-grade intelligence collection, continuous target monitoring (companies/people/topics), change detection, sentiment tracking, knowledge graph construction, critical alerts
• Predictor: Superforecasting engine, multi-source signal collection, calibrated reasoning chains, confidence-interval predictions, Brier score self-tracking, contrarian pattern detection
• Researcher: Deep autonomous research, multi-source cross-referencing, CRAAP-standard credibility assessment, APA citation reports, multilingual support
• Twitter: Autonomous Twitter/X account management, 7 rotating content formats, optimal timing scheduling, mention replies, approval queues
• Browser: Web automation agent, Playwright bridging, session persistence, mandatory purchase approval gates

Security Architecture (16-Layer Defense-in-Depth)#

WASM dual-metered sandboxing, Merkle hash chain audit trails, information flow taint tracking, Ed25519 signed agent manifests, SSRF protection, key zeroization, OFP mutual authentication, capability gating, security headers, health endpoint sanitization, subprocess sandboxing, prompt injection scanning, loop guards, session repair, path traversal protection, GCRA rate limiting.

Multi-Channel & Model Support#

40 messaging channel adapters covering IM (Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Email), enterprise (Teams, Mattermost, Google Chat, Webex, Feishu, etc.), social (LINE, Viber, Facebook Messenger, Mastodon, Bluesky, Reddit, LinkedIn, Twitch), community (IRC, XMPP, Guilded, etc.), and privacy (Threema, Nostr, Mumble, etc.). Each adapter supports per-channel model override, DM/group policies, and rate limiting.

3 native LLM drivers (Anthropic, Gemini, OpenAI-compatible) routing to 123+ models across 27 providers, with task complexity scoring for intelligent routing, automatic fallback, and cost tracking.

Access & Interaction#

• 140+ REST/WS/SSE endpoints with OpenAI-compatible API
• Web Dashboard (http://localhost:4200)
• Tauri 2.0 desktop app (system tray, notifications, global shortcuts)
• WhatsApp Web gateway (QR code connection to personal WhatsApp, no Meta Business account required)
• 53 built-in tools + MCP + A2A protocol extensions
• SQLite persistence + vector embeddings + session compaction
• FangHub: Hand/Skill marketplace
• Migration support from OpenClaw, LangChain, AutoGPT

Quick Start#

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

Configuration file at ~/.openfang/config.toml supports default model, memory decay, network listener, session compaction, channel integration, and MCP server settings. CLI provides Hands lifecycle management (activate/status/pause/list) and migration commands.

Unconfirmed Information#

• Official documentation site URL not shown in public content
• FangHub marketplace specific address not provided
• v0.5.10 release date marked as "April 2026", inconsistent with current timeline, possibly forward-looking
• Benchmark data is self-tested by the project, not third-party verified
• Specific relationship with OpenClaw project remains to be confirmed