Weft

> **Building in public, two months in.** Weft is young. The language, the type system, and the durable executor are the stable parts. The node catalog is small and intentionally opinionated (a few dozen nodes across LLM, code, communication, flow, storage, and triggers). The long-term vision is to let projects define their own nodes fluently in the language itself, but that is still ahead. If you are evaluating it for production, treat it as a foundation to build on, not a finished product. Breaking changes are expected while the shape is still settling; they will be announced, and migration notes will come with them.

>

> **Note on the docs.** This README, `DESIGN.md`, `CONTRIBUTING.md`, `CODE_OF_CONDUCT.md`, and most of the prose around the project were written fast to get the open source release out the door. They may sound a bit AI-generated in places. If you have the time and taste to rewrite any of them more cleanly, please do. A PR that improves the writing is as welcome as one that fixes a bug.

**A programming language for AI systems.**

In 2026, real software calls LLMs, spins up databases, waits for humans, browses the web, coordinates agents. Where are those primitives? You are still importing libraries and writing plumbing for things that should be one line.

Weft is a language where LLMs, humans, APIs, and infrastructure are base ingredients. You wire them together, the compiler checks the architecture, and you get a visual graph of your program automatically. No plumbing.

• **First-class humans.** Pause mid-program, send a form to a human, wait days, resume exactly where you left off. One node. No webhooks, no polling, no state management.
• **Recursively foldable.** Any group of nodes collapses into a single node with a described interface. A 100-node system still looks like 5 blocks at the top level.
• **Typed end to end.** Generics, unions, type variables, null propagation. The compiler catches missing connections, type mismatches, and brok