OpenFang

Core Concepts

Hands / Agent / Tools / Skills / Memory / Protocols

Overview
What It Is
Core Concepts
Architecture
Security (16 Layers)
Hands
Ecosystem
Getting Started
Benchmarks
Whitepaper
Sources

Core Concepts

This page breaks down OpenFang's "system primitives" by layer, helping you understand: what is an "executable unit", what handles "scheduling and governance", and what is for "extension and integration".

Hands: Autonomous Capability Packages (Core Innovation)

Hands are a set of pre-configured, long-running capability packages. You don't "chat with them to do something temporarily"; instead, you activate them. They run on a schedule, accumulate knowledge, and actively report back to the dashboard.

(The three-layer structure of Hands: HAND.toml + system prompt playbook + SKILL.md. See the "Hands" page for more details.)

Agent / Workflow / Trigger

  • Agent: An execution entity that can be spawned, receive messages, and invoke tools.
  • Workflow: A multi-step pipeline (supporting fan-out, conditionals, loops, etc.), managed by the kernel's WorkflowEngine.
  • Trigger: Event pattern matching that triggers a workflow or a background loop.

Tools / Skills

  • Tools: Capabilities provided at runtime (e.g., web fetch/search, file read/write, browser automation, image/voice processing, etc.).
  • Skills: Pluggable tool/knowledge packages. Supported types include Python/Node/WASM/PromptOnly. It also supports parsing SKILL.md (OpenClaw style) and injecting it into the context.

According to docs/architecture: 60 bundled skills are compiled into the binary using Rust's include_str!().

Channels

Channels constitute the "integration layer", connecting agents to Telegram/Discord/Slack/WhatsApp/... and handling:

  • Per-channel model/system prompt overrides
  • DM/group messaging policies
  • Rate limiting
  • Output formatting (Markdown -> platform-specific format)

Memory (SQLite Substrate)

According to docs/architecture, the memory substrate is powered by SQLite, and structured into layers including:

  1. Structured KV (per agent and shared namespace)
  2. Vector semantic retrieval (embedding + cosine similarity)
  3. Knowledge Graph (entities/relationships)
  4. Session management (preserving LLM history, token counts, crash recovery)
  5. Task board (collaboration queue)
  6. Usage & Canonical sessions (usage tracking + unifying context across channels)

Protocols: MCP / A2A / OFP

  • MCP: Model Context Protocol. OpenFang implements both an MCP client and an MCP server.
  • A2A: Google's A2A agent-to-agent task protocol.
  • OFP: OpenFang Protocol, a peer-to-peer network featuring HMAC-SHA256 mutual authentication.