Home

OpenArc

Compile typed specs into verified TypeScript with small local models.

OpenArc treats an LLM as a constrained compiler backend, not as a chat partner:

• You write the program: interfaces, classes and the members you know, with return arc.todo(); where the compiler should write a body, and tests as TypeScript (arc.test).
• OpenArc builds a context package for that one function. It's the code-generation equivalent of a unit test: only what the function needs, never the whole repository.
• A local model returns a function body.
• The compiler verifies it: type-check, static rules, then your examples in a sandbox.
• Only then is the body spliced in. Your hand edits are never overwritten.

The hypothesis behind it:

A small model working on a well-defined, isolated problem can be more reliable than a large model working on a poorly constrained, repository-wide one.

Start here

• Getting Started: install, init, compile your first region
• Code First: the main way since 0.9.0: pending bodies in functions and classes, JSDoc intent, arc.test, hidden tests, laws
• Writing Specs: code first in brief, then defineRegion, examples, blocks, sketches, concepts, TSX UI specs
• Your own palette: build your own reusable concepts: laws, variants, sketches, code-first classes and tests (the bundled palette is deprecated and frozen since 0.9.0)
• Configuration: openarc.config.json; local Ollama or APIs (OpenAI, OpenRouter, Anthropic…) per region
• Context Packages: how much the model sees, and when that grows
• Safety Model: what the model can and cannot touch, enforced in code
• Fakes: testing functions with effects
• Adopting Existing Code: openarc extract turns functions you already have into regions
• Watch Mode: openarc watch recompiles regions as you save, one at a time
• CLI Reference
• Bootstrap: OpenArc compiling OpenArc
• Design Log: every finding that changed the spec

The normative specification is SPEC-phase0.md.