This documentation site
docs.co-relayer.com is live in the repository: the guides, the agent pages, the API reference,
the contract reference and the error catalogue, with local search, llms.txt output and a link
checker that fails the build.
What is generated rather than written
Roughly two thirds of the pages on this site are not written by hand, and that is deliberate — a reference page that a person maintains is a reference page that drifts.
| Generated | From |
|---|---|
| The API reference, one page per operation | build/specs/openapi.yaml |
The plain-Markdown API summary at /api-reference.md | the same document |
| The contract's endpoints, views, events and types | corelayer.abi.json, the contract's own build output |
| One page per error code, 91 of them | the error catalogue, checked against the ErrorCode enum of the API document |
/errors.json | the same catalogue |
| The tier data behind the pricing pages | the launch pricing snapshot the marketing site renders |
| Tokens, fonts, the logo, the OpenAPI document, the ABI | @corelayer/tokens, @corelayer/ui, assets/logo, the specifications, the contract build |
A check mode regenerates everything into memory and fails on the first difference, so a stale reference page breaks the build instead of shipping.
The strongest of these is the error catalogue: an error code that exists in the API without a
page here fails this build. Since every problem document's type member is the address of its
page, that turns a machine-readable promise into something that cannot quietly rot.
For machine readers
/llms.txt— the index/llms-full.txt— every page in full, with the API reference appended/llms-agents.txt— only the agent, MCP, x402, API and error pages- any page, as Markdown — append
.mdto its URL /openapi.yaml,/errors.json,/abi/corelayer.abi.json
robots.txt allows everything and blocks no crawler, including AI crawlers. Documentation is for
reading.
The build refuses to ship a broken link
After the site is built, a script walks every page of the output and resolves every internal link and every fragment against the files that were actually produced. External links are held to an allow-list of hosts this project means to link to. It is offline on purpose: a build that reaches out over the network is not reproducible.
It also checks the agent-facing output for leaked JSX, template placeholders and unresolved snippet imports — the specific ways Markdown generated from MDX turns into noise.
What this site will not tell you
No uptime percentage, no latency target, no benchmark. CoRelayer has never run in production, so there is nothing measured to publish, and a number that is not measured is not a number. Where a page has to describe an intention it says "designed for"; where it would be guessing, it says nothing. (Status and SLOs)