ZeroClaw
Quick Facts
- Category
- Agent runtime (not a model)
- Language
- Rust
- Binary size
- 3.4MB (static, stripped)
- Startup
- Sub-10ms cold start
- Runtime RAM
- Under 5MB
- Providers
- 22 native — Anthropic, OpenAI, Bedrock, Vertex, Azure, Groq, Together, Fireworks, DeepSeek, Moonshot, Mistral, Cohere, Ollama, vLLM, etc.
- Dependencies
- Zero runtime dependencies
- Target
- Serverless (Lambda, Workers, Fly), sidecars, CLI embeds, low-spec hosts
- License
- Open source (see Claw ecosystem on GitHub)
Summary
ZeroClaw exists for the places where every constraint matters: Lambda cold starts billed in milliseconds, Cloudflare Workers capped at a few MB, Alpine-based sidecars that sit next to every microservice, CLI tools where a 200MB Node runtime is a non-starter. The engineering discipline is aggressive — no runtime reflection, no dynamic loading, no async runtime overhead it doesn't need, no features that can't be justified against the binary-size budget.
The feature set is deliberately narrow: an agent loop, a tool protocol, a provider abstraction, and basic observability. Everything else — memory, sandboxing, channel adapters, integrations — lives upstream or elsewhere in the architecture. ZeroClaw is a lean core that other runtimes (OpenClaw, MicroClaw, ExoClaw) use internally for hot paths.
Architecture
- Static Rust binary — one file, no dependencies. Works on glibc, musl, Alpine, scratch containers.
- Agent loop — minimal. Provider call → parse → dispatch tool → await → repeat. No session store, no persistence (caller brings its own).
- Provider abstraction — 22 native providers via a single trait. Switching is a config flip; failure modes route identically.
- Tool protocol — JSON-over-stdio by default; gRPC or HTTP endpoints optional.
- Observability — OTEL-compatible span emission. Everything else (metrics backend, trace storage) is the host's responsibility.
- No built-in channels, sandbox, or memory. That's the design.
When to Choose ZeroClaw
ZeroClaw is the right choice when: the deployment surface is constrained (serverless, edge, embedded); cold-start time is billed; the binary has to drop into a distroless or scratch container; or you want a consistent agent-loop implementation across many small services that each need agent capability without the operational weight of a full runtime. It's also the typical pick for CLI tools that embed an agent — terminal utilities, CI steps, local dev tools.
Tradeoffs
- Feature-sparse by design. No durable memory, no channel adapters, no sandbox, no integration library. Bring or build the rest.
- No built-in security isolation. ZeroClaw trusts its tool calls. For untrusted inputs or multi-tenant isolation, use NanoClaw.
- Configuration surface. 22 providers means 22 sets of auth, rate-limit, and error semantics to understand. The abstraction simplifies but doesn't hide.
- Not a standalone product. ZeroClaw is typically embedded in a larger system. If you want a complete deployable agent out of the box, OpenClaw or ExoClaw is the better choice.
Related
- PicoClaw — even smaller footprint; IoT and embedded.
- NanoClaw — security-first counterpart for untrusted-input workloads.
- OpenClaw — full-featured flagship that uses ZeroClaw internally.