libclaw

NOT a library or framework dependency to use for building an OpenClaw-like AI personal assistant.

But, also NOT an application to fork and customize into you own AI personal assistant.

libclaw is something in-between. It's an experiment to see if we can distribute AI instructions, code, and other assets that allow you to assemble, maintain, update, and maybe even verify your own app. And yes, libclaw is aimed at providing features that solve AI personal assistant shaped problems.

Big Picture Idea: You own your app code, libclaw knows the AI assistant features, and your agent is great at integrating the features into your app.

How does this work?

First, using libclaw depends on having an agentic coding tool to build together with. This might be Claude Code, Codex, OpenCode, Amp, or many others. This is early days, so the project is somewhat coupled to Claude Code, but we aim to use features that most of the agentic coding tools have in common. Feel free to try it out with other tools and report back what you find. We can't promise that all of the AI coding assistants will be prioritized.

Skills are the basic building block of libclaw. A skill can package up both instructions and code. Instructions help us leverage the power of models and agents. Code provides specificity. Skills are relatively token efficient (agents are decent at finding the right skill at the right time).

Maybe skills are all we need. Maybe not. Let's figure it out together.

Features

Starting with a small, focussed set of features:

1. Agent in a box: An agentic runtime in an isolated context
2. Human to Agent bridges: An example is a chat based app like Slack to the agent in the box. But more generally, it doesn't have to be chat based, it could be a UI, voice, or something else.
3. Data bridiges: An example is an MCP server used to get data to and from Gmail. But this can also be generalized.

Technology choices

We're trying to support a few JavaScript runtimes: Deno, Node.js, and Cloudflare Workers. You should be able to leverage libclaw without feeling limited to a specific deployment target. For now, Cloud based deployments are more interesting than deployments on a personal computer.

Isolation is provided by Docker containers. You probably don't want the agent messing with the system that's running it. It's also nice to be able to manage separate context windows, and force all sharing to be explicit.