What It Is (Positioning & Design Philosophy)
The OpenFang README explicitly emphasizes that it is not a chatbot framework, not a Python wrapper, and not merely a multi-agent orchestrator. Instead, it is a production-grade Agent Operating System.
Traditional Frameworks vs Agent OS
Where Is the "More Radical, Kernel-Level" Approach Manifested?
As seen in docs/architecture, OpenFang pushes many capabilities traditionally treated as "business logic" down into system-level primitives:
- Kernel Boot Sequence: Like an operating system, it has a rigid boot sequence (loading configs, migrating memory, assembling subsystems, restoring persistent agents, connecting to MCP...).
- Capability Management: Transforms "can it call a specific tool?", "can it connect to a host?", or "can it spawn/kill an agent?" into declarative, verifiable, and inheritable constraint sets.
- WASM Dual Metering: Tool execution defaults to a Wasmtime sandbox, balancing a "deterministic instruction budget (fuel)" with a "wall-clock time budget (epoch interrupt)".
- Multi-Layer Security: Mechanisms like SSRF protection, path traversal protection, prompt injection defense, secret zeroization, rate limiting, and health redaction are built as independent systems.
Relationship with OpenClaw (Briefly)
Both the official website and the README mention that OpenFang is inspired by OpenClaw. It provides a migration tool (openfang-migrate) to translate OpenClaw YAML to OpenFang TOML, while offering compatible tool name mapping and SKILL.md parsing.