# MCP

okmodel does not ship an MCP server. There is no package to add, no config file, and no tools to call.

An MCP server is deferred. If one is built, it is a separate package outside the core. It is not scheduled.

## In the browser

The site registers three read-only tools when the browser has `document.modelContext` or `navigator.modelContext`. A browser without that API does not download the script.

| Tool | Input | Reads |
| --- | --- | --- |
| `search_docs` | `query` | `/api/search` |
| `get_page` | `slug`, such as `reference/cli` | That page as markdown |
| `list_pages` | none | `/llms.json` |

## The handbook

An agent reads the site. Start at [`/llms.txt`](/llms.txt). The index lists every public page as markdown, in sidebar order. Each line is the page title, its markdown URL, and the page description.

| URL | Returns |
| --- | --- |
| `/llms.txt` | The index, as markdown. |
| `/llms.mdx/docs/⟨slug⟩.md` | One page. The title is the heading. YAML frontmatter is stripped. |
| `/llms-full.txt` | Every page in one file. Each block starts with the title and the HTML path. |
| `/llms.json` | The same pages as JSON. Each page has `slug`, `title`, `description`, `html`, and `markdown`. |

The introduction is `/llms.mdx/docs/index.md`. A nested page keeps its path, such as `/llms.mdx/docs/define/table/columns.md`. The same route also accepts a trailing `content.md` and the slug without `.md`.

Send `Accept: text/markdown` on a handbook URL and the response is that page's markdown. The response sets `Vary: Accept`. This covers `/`, `/docs`, and `/changelog`. A URL under `/llms` is already markdown and is served as written. An unknown path with that `Accept` header is a markdown 404. The body links `/docs`, `/llms.txt`, and `/sitemap.xml`.

## Errors

`OkmError`, from `okmodel`, carries `fix` for a person or a coding agent. `fix.summary` is one sentence. `fix.suggestion` is the nearest accepted name, when one is close.