Index / ZeptoClaw · updated Oct 24, 2026
ZeptoClaw
qhkm/zeptoclaw · healthy · rank 36 of 50 by stars
A single ~6MB Rust binary that packs tools, memory, channels, 18 providers, and sandboxed autonomy into a local-first assistant runtime. It positions itself as the synthesis of OpenClaw's ecosystem, NanoClaw's isolation, NemoClaw's guardrails, and PicoClaw's footprint discipline.
Facts
Repository
Runtime
Posture
Community
Security breakdown
Composite 82 / 100 · how these are scored
higher is safer
higher is safer
higher is safer
higher is safer
higher is riskier
Model access
Read from the repository, not written by a model · 54 files examined
5 providers · gateway support
Pinned models
gpt-4o released 2024-05-13 — 2 years old (from the public model catalogue)
Pin last edited 7 months ago
Several defaults are pinned; the project picks one per provider rather than shipping a single default.
Evidence
Decision
Why choose ZeptoClaw over OpenClaw?
Why choose this
- ~6MB binary and ~6MB RAM versus OpenClaw's large TypeScript footprint
- ~50ms startup suitable for edge and $10-class hardware
- Built-in container isolation, prompt-injection detection, and circuit-breaker provider stack
Tradeoffs
- Far smaller community, integrations, and skills ecosystem
- No Reddit presence and mixed ecosystem-digest sentiment labeling it low-velocity
- CI checks were removed from GitHub Actions, weakening automated merge gating
Best fit
- Users wanting a tiny single-binary agent runtime on edge or VPS hardware
- Operators running local reasoning models via Ollama or LiteLLM
- Security-conscious users who want container isolation and prompt-injection detection without Docker/k3s overhead
Avoid if
- You need a mature plugin/skills ecosystem and large community
- You require enforced CI gates — the project removed GitHub Actions CI checks
- You want production-proven stability; ecosystem digests flag it as low-velocity
README and commit history give strong evidence for architecture, performance claims, and security posture. Community sentiment is thin (zero Reddit matches) and web digests conflict with the active commit cadence, so adoption-risk fields carry moderate uncertainty.
AI layer reviewed Oct 5, 2026 · how this is written
Star activity
650 stars today
Overview
ZeptoClaw is a personal AI assistant infrastructure written in Rust, shipped as a single ~6MB binary with ~50ms startup and ~6MB RAM usage. It bundles tools, long-term memory, messaging channels, 18 model providers, and sandboxed autonomy, explicitly positioning itself as a synthesis of lessons from OpenClaw (integrations), NanoClaw (container isolation), NemoClaw (policy-gated guardrails), and PicoClaw (edge efficiency) without their respective tradeoffs.
Architecturally, it emphasizes security and correctness at the edges: container isolation, prompt-injection detection, a circuit-breaker provider stack, tool-schema sanitization for strict providers (Anthropic) and local grammar-constrained backends (llama.cpp), and conservative coercion of malformed tool arguments from small local models. Recent work shows deep attention to the local-model deployment story — parsing reasoning_content from OpenAI-compatible endpoints so reasoning models like Qwen3 and DeepSeek distills never return silent empty replies, with explicit guards preventing chain-of-thought leakage into tool calls or streams.
Compared to OpenClaw, ZeptoClaw trades ecosystem breadth for footprint and control: no large TypeScript app, no heavy config surface, and a local-first posture with optional cloud providers. The project maintains a large test suite (3,900+ tests claimed) and active dependency security hygiene (prompt Rustls advisory patching), though it recently removed GitHub Actions CI in favor of local validation gates, and third-party ecosystem digests characterize it as lower-velocity than the leading clones.