https://github.com/nullclaw/nullclaw

I kept running into the same problem with AI agents: they work, but they’re heavy. Most frameworks assume Python or Node, pull in hundreds of MB of dependencies, and feel overkill if you just want an always-on assistant that starts instantly and runs anywhere.

So We built NullClaw.

NullClaw is a fully autonomous AI assistant infrastructure written in Zig. It compiles to a static binary (~678 KB), uses about ~1 MB of RAM at runtime, and starts in a couple of milliseconds. No Python, no Node, no containers required.

What surprised me most while building it wasn’t just the size, but how much functionality you can keep without the usual bloat:

Supports 20+ AI providers

Multiple channels (Telegram, Discord, Slack, etc.)

Tool execution with sandboxing

Hybrid memory (vector + full-text search)

Secure secrets handling and scoped workspaces

Pluggable architecture for providers, tools, memory, and tunnels

The idea is simple: an AI assistant should behave more like a system daemon than a web app.

NullClaw is designed to run comfortably on:

low-cost VPSes

edge devices / SBCs

local machines as an always-on agent

or as a lightweight backend for larger agent systems (it’s compatible with OpenClaw)

If you want a higher-level overview, examples, and use cases, I’ve put that together at https://nullclaw.co/ — the GitHub repo stays focused on implementation details.

This project is also an experiment in whether compiled languages make more sense for agent infrastructure than interpreted stacks. Zig gave me predictable performance, strong safety guarantees, and binaries that are genuinely easy to ship.

I’m curious how others here think about:

agent runtimes vs frameworks

edge / local-first assistants

and whether we’ve collectively accepted too much infrastructure overhead as “normal”

Happy to answer questions or hear criticism.