Developers
How AI agents and LLM tools read the CentraKit docs through llms.txt, Markdown page URLs, the docs MCP server and Ask AI, and how they act on CentraKit data through the product MCP server.
Every page of these docs is available as Markdown, and the site has an index, a full-text export and an MCP server for agents. Use this page to point an assistant, a coding agent or a retrieval pipeline at the docs, and to choose between reading docs and acting on CentraKit data.
| You want to | Use |
|---|---|
| Give an agent a map of the docs | https://docs.centrakit.com/llms.txt |
| Load all docs into one context | https://docs.centrakit.com/llms-full.txt |
| Fetch one page as Markdown | The page URL with .md, or Accept: text/markdown |
| Let an agent search and read docs on demand | The docs MCP server at https://docs.centrakit.com/mcp |
| Ask a question in the browser | Ask AI on any docs page |
| Read or change customers, tasks, quotes and orders | The product MCP server at https://api.centrakit.com/api/mcp |
/llms.txt is a plain-text index of every page with its title, URL and description. It is small enough to put in a system prompt, and agents use it to pick the pages to fetch.
/llms-full.txt is the Markdown of every page in one file. It is large, so prefer the index plus single pages unless you are building a retrieval index or the model has a large context window.
Both cover the guides, the OpenAPI operation pages and the changelogs.
Every page has a Markdown version. Add .md or .mdx to the page path:
curl https://docs.centrakit.com/developers/mcp.md
curl https://docs.centrakit.com/developers/mcp/index.mdxOr ask for Markdown on the normal URL with the Accept header:
curl -H "Accept: text/markdown" https://docs.centrakit.com/developers/mcpWhen a request prefers text/markdown, the docs site returns the Markdown instead of HTML and sets Vary: Accept, so caches keep both versions apart. Responses use Content-Type: text/markdown; charset=utf-8.
Each Markdown page starts with # <title> (<url>) and the page description, followed by the page body. OpenAPI operation pages contain only that page's operations and the component schemas they reference, not the whole specification.
Every docs page has two buttons under its title:
The docs MCP server lets an agent search and read the docs as tools instead of fetching URLs. It is public, read-only and needs no authentication.
| Environment | Endpoint |
|---|---|
| Production | https://docs.centrakit.com/mcp |
| Local | https://docs.localhost/mcp |
It has three tools:
| Tool | Input | Returns |
|---|---|---|
list_pages | none | The page index, the same content as /llms.txt. |
search | query, optional locale | Matching sections with their page URLs. |
get_page | url, a page pathname such as /developers/mcp | The page as Markdown, or page not found for an unknown URL. |
Add it to a client that supports remote MCP servers:
{
"mcpServers": {
"centrakit-docs": {
"url": "https://docs.centrakit.com/mcp"
}
}
}For one-click installs in Cursor, VS Code and Claude Code, see the connect table on MCP server.
The floating Ask AI button on every docs page opens a chat panel. The assistant searches the docs with a search tool, answers from the results, and cites the pages it used as links. When the docs do not cover a question, it says so and suggests a better search.
Ask AI uses the platform's default chat model, which administrators set in AI settings. It reads only the public docs and cannot see your organization data.
To let an agent act on your CentraKit data, connect the product MCP server at https://api.centrakit.com/api/mcp. It needs OAuth, runs with your permissions, and exposes 74 tools for tasks, customers, products, quotes, commerce, agenda, files, workflows, webhooks and the audit log. Mutations can be held by approval policies. See MCP server.
A typical setup connects both servers: the docs server so the agent knows how CentraKit works, and the product server so it can do the work.
The repository has instructions for coding agents that work on CentraKit itself:
AGENTS.md at the repository root describes the layout, commands, architecture invariants, conventions and hard rules. CLAUDE.md points to it.docs/agents/ holds notes on mistakes agents have repeated, one page per topic: the toolchain, git and release, the Supabase schema, tenancy and routing, translations, testing and writing docs. Agents load the page for the topic they are working on..agents/rules/ and .agents/skills/ hold rules and skills that attach to specific files or tasks.Docs pages are written from the shipped code. Before you change a page in this site, read Writing docs.
Last updated on