Specs, not packages
A registry of machine-readable library specifications. AI agents fetch a spec and generate verified native implementations in any language — no dependency, no supply chain.
How it works
A spec is the contract. The implementation is generated and verified — in whatever language you ship.
Fetch a spec
Pull a machine-readable specification — types, functions, behavioral contracts — over a plain HTTP API.
Generate native code
An agent reads the spec and writes a first-party implementation in your language. No third-party package to trust.
Verify with test vectors
Every spec ships test vectors. Run them against the generated code to prove it matches the contract.
Why Canra
At first it seems backwards: why have every project regenerate code instead of reusing a package? Because generating from a verified spec buys you things a dependency can't.
No supply chain risk
Nothing runs from a third party. The implementation is generated and reviewed in your own tree — fully auditable, no opaque transitive dependencies.
Minimal footprint
You generate only what you actually use, instead of pulling in a whole library — and everything it depends on — for one function.
No licensing concerns
Generated from an open spec, the code is yours. No package licenses to track, audit, or comply with across your dependency tree.
AI agents make writing code cheap — and each spec's detailed contracts and test vectors keep what they write correct. It works with any language and any agent, and it's a natural fit for AI-first toolchains like Maxon.
# 1. fetch the spec + test vectors
$ curl https://canra.dev/api/specs/canra/iso-8601
# 2. generate a native implementation, any language
$ agent generate iso-8601 --lang rust
# 3. verify against the spec's test vectors
$ agent verify iso-8601
42 vectors passed — implementation matches the contractPublish your spec
Define a library once, as a verifiable spec. Let every language — and every agent — implement it natively.