Started as an ideation before ChatGPT, when models couldn't even produce a valid JSON yet. A DSL that lets an LLM compose an entire web and mobile app — one parse tree to Next.js and Expo — from a syntax simple enough that designers wrote it too, and constrained enough that a model can't get it wrong.

Better primitives for humans and agents working on Next.js.

A shared store of auto-optimized, auto-evaluated software snippets. An agent searches for an existing snippet that performs the action it needs, runs it, and publishes suggested changes and evolutions back — so every agent in the world can contribute to improving them.

Built before Anthropic's Agent Skills. The domain has since been sold; the work is being republished as OpenIntegrations and MuscleMemory (soon).

A harness that lets an agent modify its own code and spread across machines. It writes its own tools for whatever platform it lands on, and gains pervasive capabilities from an ever-growing mesh of devices it runs on.

Try — create an agent, run the command on your own machine, then ask it to photograph you by building the tool first, to spawn itself a server with the AWS CLI, or to spread to a Raspberry Pi on your network.

A way for agents to identify themselves without creating accounts. Identity is a keypair, not a name; a service checks one key and learns which machine is calling, whether a human stands behind it, and what that human allowed.

Lets agents contact any web API in the world — 6,000+ providers — without ever seeing the credentials. The agent sends a placeholder; the proxy injects the real token. Nothing to steal from a context window.

Account-less resource provisioning for agents. npx extraorbital provision . reads the project, creates the databases, storage and model keys it needs, and writes the credentials to .env.local. No console, no card, safe to retry.