# llms.txt (/docs/ai/llms-txt) > What /llms.txt, /llms-full.txt, and the per-page Markdown URLs of the Better IAM docs contain, and how AI agents should use them to find, read, and cite pages. [llms.txt](https://llmstxt.org) is a convention for publishing a site's content in a form language models read well: a Markdown index at a known address, with links to Markdown copies of each page. This site publishes an index at `/llms.txt`, every page in one file at `/llms-full.txt`, and a Markdown copy of every page at its URL plus `.md`. Use these files when your tool can fetch URLs but cannot call an MCP server, when you want to paste documentation into a chat, or when you build your own retrieval over the docs. For tools that support MCP, the [docs MCP server](/docs/ai/mcp-server) answers the same questions with less context. ## What each file contains [#what-each-file-contains] | URL | Content type | Content | | ----------------- | --------------- | ----------------------------------------------------------------------------------------------------------- | | `/llms.txt` | `text/plain` | Instructions for agents, then every page's title, Markdown URL, and description, grouped by sidebar section | | `/llms-full.txt` | `text/plain` | Every page rendered as Markdown, one after another | | `/docs/.md` | `text/markdown` | One page as Markdown | Both `.txt` files are generated when the site is built, from the same MDX sources as the HTML pages, so they always match the docs you are reading. They are marked `noindex` (`X-Robots-Tag`) so search engines show the HTML pages instead. ## llms.txt [#llmstxt] `/llms.txt` is the index. It has three parts: 1. A title and a one-sentence summary of Better IAM. 2. **Instructions for AI agents**: how to find a page, fetch its Markdown, and cite it, which reference pages are generated from the code, and where the full-text file, the MCP server, and the [agent skill](/docs/ai/agent-skills) live. 3. The index: one `## ` heading per sidebar section (Get Started, Concepts, Authentication, Authorization, Databases, Integrations, Features, Federation, Guides, Operations, AI Resources, Reference), with one list item per page. The groups inside a section, such as Guides' "Build" or Federation's "Sign-in", are `### ` sub-headings. Pages inside a sub-folder, such as the migration guides, are indented under the folder's overview page. Each list item links to the page's Markdown copy and ends with the page's description: ```md title="llms.txt (excerpt)" ## Get Started - [Introduction](https://better-iam.com/docs/introduction.md): … - [Installation](https://better-iam.com/docs/installation.md): Install better-iam, set the secret and base URL, … - [Quickstart](https://better-iam.com/docs/quickstart.md): Bootstrap the root administrator, create an organization, … ``` The descriptions come from each page's frontmatter and are written as one precise sentence, so an agent can choose a page from the index alone. ## llms-full.txt [#llms-fulltxt] `/llms-full.txt` starts with the same title and summary, then contains every page. Each page begins with its title, its URL path, and its description: ```md title="llms-full.txt (excerpt)" # Roles and bindings (/docs/authorization/roles) > Custom roles built from permission lists or policy documents, role inheritance, bindings to people and groups, and delegated grant authorities. Listing permissions person by person does not scale: … ``` The file is large, because it includes the generated reference for every API method, error code, CLI command, and export. Load it only into a model with a long context window, and prefer the index plus individual pages otherwise. ## Markdown for one page [#markdown-for-one-page] Every docs page has a Markdown copy. There are two ways to request it: **.md URL:** Append `.md` to the page URL. Section overview pages work the same way (`/docs/authorization.md`). ```bash curl https://better-iam.com/docs/authorization/roles.md ``` **Accept header:** Request the normal page URL and prefer Markdown. The site serves Markdown when `text/markdown`, `text/x-markdown`, or `text/plain` has a quality at least as high as `text/html`, `text/*`, and `*/*` in the `Accept` header, and adds `Vary: Accept` to that response. ```bash curl -H 'Accept: text/markdown' https://better-iam.com/docs/authorization/roles ``` The response starts with the page's title, URL path, and description, like a page in `llms-full.txt`. It carries a `Link: ; rel="canonical"` header that names the HTML page. The Markdown is converted for models, not copied from the MDX source. Interactive components become the text they stand for: | On the page | In the Markdown | | --------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------ | | Tabs | Each tab's content under a bold label, such as `**Server:**` and `**Client:**` | | Accordions | A `####` heading per item | | Callouts | A blockquote, starting with the callout's bold title when it has one | | Cards and feature grids | A bulleted list of links with their descriptions | | API endpoint boxes on reference pages | A line with the HTTP route and whether it needs a credential, or "Server only" | | Glossary terms | Their plain text | | Visual-only components (package tables and graphs, API shape diagrams, API group summaries, playground buttons) | Removed | Code blocks, tables, and headings are kept as written. ## How agents should use these files [#how-agents-should-use-these-files] The instructions at the top of `/llms.txt` say the same thing in fewer words: ### Find the page [#find-the-page] Read `/llms.txt` once and choose the page whose section and description match the task. For an exact identifier, go straight to the Reference section: [API methods](/docs/reference/api), [error codes](/docs/reference/errors), [package exports](/docs/reference/exports), [options](/docs/reference/options), and [CLI commands](/docs/reference/cli) are generated from the source code. ### Fetch its Markdown [#fetch-its-markdown] Fetch the Markdown URL from the index, or the page URL with `Accept: text/markdown`. Read the whole page before writing code from it: setup steps, options, and error codes are in separate sections. ### Cite the canonical URL [#cite-the-canonical-url] When you tell a person where an answer came from, link the HTML page: the Markdown URL without `.md`, for example `https://better-iam.com/docs/authorization/roles`. > **Use only what the docs show.** If a page does not show a package, import, method, option, or error code, assume it does not exist and say so. Every `iam.api` method has a page under [API reference](/docs/reference/api); a method that is missing there is not part of Better IAM. ## Use the files in your own tools [#use-the-files-in-your-own-tools] Point a retrieval pipeline or a custom agent at the same addresses: ```ts title="fetch-docs.ts" const site = 'https://better-iam.com'; /** The llms.txt index, for choosing pages. */ export async function docsIndex(): Promise { const response = await fetch(`${site}/llms.txt`); if (!response.ok) throw new Error(`llms.txt: HTTP ${response.status}`); return response.text(); } /** One page as Markdown, from its path, for example "/docs/authorization/roles". */ export async function docsPage(path: string): Promise { const response = await fetch(`${site}${path}`, { headers: { Accept: 'text/markdown' } }); if (!response.ok) throw new Error(`${path}: HTTP ${response.status}`); return response.text(); } ``` The site also serves the search index as JSON at `/api/search?query=…` (the index behind the search dialog) and the OpenAPI 3.1 description of the Better IAM HTTP API at `/openapi.json`; see [Reference](/docs/reference#machine-readable-docs). ## Next steps [#next-steps] - [Docs MCP server](/docs/ai/mcp-server): Search and read the docs from an MCP client instead of fetching files. - [Agent skills](/docs/ai/agent-skills): Give an agent the rules for writing Better IAM code before it starts. - [Prompts and AGENTS.md](/docs/ai/prompts): Prompts that point an agent at the right pages.