Weft treats humans, LLMs, and Postgres as language primitives

What it does

Weft is a domain-specific language for building AI systems. You write programs that wire together LLM calls, HTTP requests, databases, and human-in-the-loop steps; the compiler validates types and architecture upfront, and emits a visual graph automatically. Execution is durable by default, backed by Restate, so a crash mid-workflow resumes exactly where it left off—even if that means waiting days for someone to fill out a form.

The interesting bit

The language collapses the boundary between code and visual workflow. Every program has two native views: dense text for editing and a node graph for reading, and edits in either sync to the other. Groups of nodes recursively fold into single nodes with described interfaces, so a hundred-node system can look like five blocks at the top level.

Key highlights

• Typed end to end with generics, unions, and null propagation; the compiler catches missing connections before runtime.
• Human queries are first-class citizens: one node pauses execution, sends a form, and resumes days later without webhooks or polling logic.
• Built-in node catalog covers LLM inference, code execution, messaging (Discord, Slack, Telegram, WhatsApp, X, Email), web search, Postgres, and infrastructure provisioning.
• Durable execution via Restate means restarts and crashes are invisible to the program logic.
• The catalog/ directory is the source of truth: each node is a Rust backend file and a TypeScript frontend definition, symlinked into the compiler and dashboard.

Caveats

• The project is two months old and explicitly “young”; the authors warn of expected breaking changes and a small, opinionated node catalog (a few dozen nodes).
• Custom node definitions written fluently in the language itself are a long-term vision, not a current feature.
• The documentation and prose were written quickly for the open-source release and, by the authors’ own admission, “may sound a bit AI-generated in places.”

Verdict

Worth exploring if you are building complex agentic workflows and want type safety plus durability without gluing together separate tools. Skip it if you need a mature, stable language or a large ecosystem of third-party nodes today.