# Use the docs with AI (/docs/reference/ai) > Connect AI coding assistants to these docs through their MCP endpoint, or hand them the site as Markdown with llms.txt and per-page Markdown. AI coding assistants write better Better Compute code when they can read the real reference instead of guessing method names, IAM actions or CLI flags. This site offers the same content to tools as it does to people, in three forms: an MCP server an assistant can query while it works, plain-text indexes for one-shot context, and a Markdown version of every page. > **Two different MCP servers.** This page is about the **documentation's** MCP endpoint, which answers questions about Better Compute. To let an agent act on a deployment (list pools, reserve GPUs, create instances), use the `@better-compute/mcp` package instead; see [MCP](/docs/guides/clients/mcp). ## Connect an assistant with MCP [#connect-an-assistant-with-mcp] The [Model Context Protocol](https://modelcontextprotocol.io) lets an assistant call tools while it works. The docs serve an MCP endpoint at `/api/mcp` over Streamable HTTP. It needs no sign-in and only reads. The server introduces itself as `better-compute-docs` and offers six tools: | Tool | Input | What it returns | | -------------------- | -------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `search` | `query`, optional `locale` | Full-text search over every page, the same index as the search dialog, as JSON. | | `list_pages` | none | Every page with its URL and one-line description (the body of `/llms.txt`). | | `get_page` | `url` | One page as Markdown, by its path, for example `/docs/guides/capacity/reservations`. | | `lookup_api_method` | `method` | One API method: its HTTP route, the server call `compute.api.{group}.{method}(credential, input)`, the browser client call, its kind, the IAM action and resource it is decided on, its module, and whether this release serves it or it is planned for a later phase. | | `lookup_error_code` | `code` | What a `ComputeError` code means and its HTTP status, for example `CAPACITY_UNAVAILABLE` or `ACCESS_DENIED`. | | `lookup_cli_command` | `command` | Usage, flags (with defaults and environment variables) and examples of one `better-compute` command, for example `dev` or `fleet apply`. | The lookup tools are forgiving about how you name things. `lookup_api_method` accepts `reservations.create`, `compute.api.reservations.create(...)`, `client.joinTokens.create`, a route such as `/api/compute/reservations/create`, or a reference URL such as `/docs/reference/api/reservations#create`; an unknown group or method answers with the list of valid ones. `lookup_cli_command` drops a leading `better-compute` and any flags, and matches the longest command name the query starts with, so `fleet plan fleet.json` finds `fleet plan`. Error codes are matched case-insensitively. **Claude Code:** ```bash claude mcp add --transport http better-compute-docs https://better-compute.com/api/mcp ``` **Cursor:** ```json title=".cursor/mcp.json" { "mcpServers": { "better-compute-docs": { "url": "https://better-compute.com/api/mcp" } } } ``` **VS Code:** ```json title=".vscode/mcp.json" { "servers": { "better-compute-docs": { "type": "http", "url": "https://better-compute.com/api/mcp" } } } ``` These docs are served at `https://better-compute.com`. When you run them locally with `pnpm --filter @better-compute/docs dev`, use `http://localhost:4100` instead. The links the lookup tools print use the site's canonical origin, `DOCS_SITE_URL`, which defaults to `https://better-compute.com` in a production build and to the local address under `next dev`. A session might look like this: the assistant searches for "extend a reservation", reads `/docs/guides/capacity/reservations` with `get_page`, then calls `lookup_api_method` with `reservations.extend` to get the exact route and IAM action before it writes the call. ## Give a model the whole site [#give-a-model-the-whole-site] When a tool cannot call MCP, paste or fetch the site as plain text: | URL | Content | Use it for | | ----------------- | -------------------------------------------------------------------------------------------------- | --------------------------------------------- | | `/llms.txt` | An index: every page's title, URL and description, under a one-paragraph summary of Better Compute | Letting a model pick which pages to read | | `/llms-full.txt` | Every page as Markdown in one file | Loading everything into a long context window | | `/docs/.md` | One page as Markdown | Quoting a single page precisely | Any docs URL also returns Markdown when the request asks for it with `Accept: text/markdown`, so agents that fetch pages get Markdown without special URLs: ```bash curl -H 'Accept: text/markdown' https://better-compute.com/docs/guides/quickstart curl https://better-compute.com/docs/guides/quickstart.md # the same answer ``` Both are rewritten to `/llms.mdx/docs//content.md`, which answers with a `Link: rel="canonical"` header that points at the HTML page. Each page's **Copy Markdown** button and its view-options menu use the same Markdown. The Markdown is not the raw MDX source. Each page starts with its title, URL and description, comments are dropped, visual-only components are removed, the API endpoint boxes become a line of text with the route, authorization and calls, and the [glossary](/docs/reference/glossary) becomes every term as a heading with its definition. `/llms.txt` and `/llms-full.txt` carry `X-Robots-Tag: noindex`, so search engines index the pages themselves. ## Keep answers accurate [#keep-answers-accurate] The reference pages, and the MCP lookup tools, come from the code: the generator builds a control plane with every module and reads its method registry, command kinds, error codes and CLI manifest (see [reference](/docs/reference)). They match the version of Better Compute these docs were built from, including which methods that version serves and which PLAN.md lists for a later phase. Two things to tell an assistant: * **Planned is not served.** `lookup_api_method` says `Status: planned for ` for methods this release does not route; the HTTP endpoint answers `NOT_FOUND` for them. Check the status before generating code against a method. * **The daemon is not released.** `bcd` is a Go binary whose v0.1 node agent (with tunnels) is built and tested end to end, and whose site and ramdisk modes are written with Go unit tests, but no binary is published. Its VPS host runs instances on `libvirt-kvm` or `cloud-hypervisor` (both tested only against fake tools, never on real hardware) or, on fake-hardware builds, a file-backed runtime; its sim mode (many node agents in one process) runs only on fake-hardware builds; self-update and the WebSocket fallback are written, and GPUs are read through `nvidia-smi` rather than the NVML library (cb987e5). The TypeScript `FakeDaemon` in `@better-compute/testing` runs those paths. See [what bcd does today](/docs/daemon#what-bcd-does-today). When an answer disagrees with your installed version, trust the TypeScript types in your editor: every server method is typed end to end from the `betterCompute()` instance and the `@better-compute/client` client.