Some of your readers are not people.
A developer pairing with an assistant needs the documentation in a shape the assistant can swallow. ApiGlow generates those files from the loaded schema, client-side — it calls no model and runs no server of its own.
The hand-off, drawn — the page as prose, the same page as Markdown, and the files an agent takes away.
The two files an assistant asks for.
llms-full.txt is the whole documentation as one Markdown file: the API info, the servers, your prose pages, every operation and every webhook. It is what you paste into a context window when you want the assistant to simply know the API.
llms.txt is the index of the same content, following the llmstxt.org convention — one link per operation and webhook, pointing into the page’s own routes. It is what an agent fetches when it would rather navigate than swallow.
Both are produced from the schema in the reader’s browser. No model is called, and nothing is uploaded to make them.
A hand-off menu on every page.
Any page can be copied as Markdown or viewed as Markdown, so an agent gets the prose without the chrome. The header’s Copy for AI menu, the same on every page, hands over the whole API: both text files, a claude mcp add one-liner, and install links for Cursor and VS Code. The home page carries the same registration as a JSON block.
The MCP card wires one of two off-the-shelf OpenAPI-to-MCP bridges — your pick — and comes out with credentials as placeholders rather than values. Where a security scheme is not a header, the card says so plainly instead of generating something that will not work.
And an agent that fixes the schema.
The schema audit has an agent readiness category: what an agent gets when it calls your API through a tool. Inputs with no type, unions it cannot tell apart, enum values nothing explains, errors returned as prose — each finding names what concretely goes wrong in an MCP bridge or a GPT Action.
The npm package ships an agent skill for the other direction. It runs apiglow audit, reads the JSON report — every finding placed at file:line:column, with its fix — corrects the schema and audits again. It never invents what your API does to fill a description, and never makes the run pass by switching a rule off: those stay your decisions.
And the crawlers that run no JavaScript.
Assistants that fetch rather than browse have the same problem search engines do: a single-page app is one URL. The static mirror is the answer —apiglow bake writes an HTML and Markdown page per route, a sitemap, and a served llms.txt, from a command you run beside your deploy.What it is not.
ApiGlow does not run an assistant. There is no chat in the documentation, no model behind it, and no key to supply — it produces files, and something else reads them.
The MCP card configures a third-party bridge. What that bridge supports is the bridge’s business, and the card is honest about the cases it cannot cover.