Early days, building in public. Weft is two months old. I am shipping fast and learning by releasing. These docs were written in a sprint against the open source launch, with heavy AI assistance, and audited against the real codebase line by line. Most screenshots are not in yet; the placeholders describe what should be there and are easy enough to find by poking around the Builder. The rest will fill in over the coming days.
A few things to calibrate your expectations:
- The current catalog is small. You can build real things with it (LLMs, human in the loop, webhooks, Postgres durable storage, WhatsApp, Slack, Discord, and a few others), but the vision is much bigger than what ships today. Today you compose from the preexisting catalog; tomorrow you will be able to define your own nodes fluently as part of writing a project, without leaving the language.
- The language, the type system, and the durable executor are the stable parts. Node contributions, tooling, and UI polish are where the churn is.
- This is beta. If you run a real workload on it, plan for rough edges. I am pushing hard to close that gap fast.
- Breaking changes are expected while the shape is still settling. They will be announced, and migration notes will come with them.
If any of this excites you, the fastest way in is Discord. Come say hi, tell me what you are trying to build, file issues when things break, argue with the design. Early contributors shape the project more than late ones. If you find anything in the docs that does not match what the code actually does, file an issue.
Weft
A programming language for AI systems.
Two native views
Weft has two views of the same program. The code view is optimized for AI: dense, structured, fast to read and write. The graph view is optimized for humans: visual, zoomable, click to inspect any node. Edit either one, the other updates instantly. The AI writes code. You see the architecture.
Loading playground...
Live playground. Click on nodes, zoom, pan.
Recursively abstractable
AI systems grow fast. Without a way to compress complexity, everything becomes unmanageable. In Weft, any piece of logic can become a block with a clear interface. Blocks nest inside blocks. Collapse to see the big picture, expand to see the details. That's how humans stay in control, and how AI agents will work on different parts independently.
Loading playground...
Try collapsing and expanding the groups in the graph to see how complexity compresses.
AI, humans, and services are first-class citizens
Most languages were designed for math and logic. LLMs, APIs, human decisions? Bolted on with libraries. In Weft, these are the primitives. An LLM call, a human approval, a database: all typed blocks that you connect like building blocks. Some are pre-built. Some you write yourself in any language. Weft is the coordination layer that makes sure everything fits together.
Loading playground...
If it compiles, it runs
The compiler catches the bugs: wrong types, missing connections, broken logic. If your code compiles, it works. At runtime, when something goes wrong (a service fails, data is missing), the system handles it gracefully instead of crashing. The only failures are external. Your logic is guaranteed correct.
Loading playground...