ZeroClaw
A Rust-based AI agent runtime with a 3.4MB binary, under 5MB RAM at runtime, and sub-10ms startup. Built by students from Harvard, MIT, and Sundai.Club. Its trait-driven architecture lets you swap providers, channels, tools, and memory backends without touching core code.
Key Facts
- Language: Rust
- Footprint: 3.4MB binary, under 5MB RAM, under 10ms startup
- Architecture: Trait-driven; swappable Provider, Channel, Tool, Memory, Runtime, Security, Tunnel
- LLM providers: 22+ with OpenAI-compatible endpoints
- Security: Workspace-only filesystem scope by default; explicit command allowlist; forbidden paths (.ssh, .aws, .gnupg); pairing required
- Memory: SQLite-centric hybrid retrieval (vector + keyword)
- License: Apache-2.0 / MIT
Best For
Developers who need an always-on agent on a cheap VPS, home lab, or Raspberry Pi where stability and low resource consumption matter more than a large skill marketplace. Teams that want security-by-default rather than security-by-configuration.
Note: Requires a Rust toolchain to compile (around 1GB RAM for build). The plugin ecosystem is smaller than OpenClaw's. Docker and WASM runtime support are planned but not yet merged.
Compared to OpenClaw
OpenClaw idles at roughly 400MB; ZeroClaw uses under 8MB. On shared or edge hardware, that difference can determine whether a project is viable. ZeroClaw's security defaults—workspace scope, allowlists, pairing—are built in rather than bolted on, which addresses concerns raised about OpenClaw's plugin ecosystem.