MCP Documentation Server#

kinetis/mcp-docs owns the catalogue of this documentation and the fetch behind it, and serves every page as an MCP resource — plus a tool that returns a bounded line window of a page and one that finds the lines of a page containing a literal string — so a coding agent reads Kinetis’s current documentation instead of answering from training data. It is framework-agnostic: kinetis/mcp-protocol is the only Kinetis package it depends on — not the framework, and not kinetis/mcp — and it knows nothing about your code. To let an agent call your own application, see Model Context Protocol (MCP).

Note

A project that already registers Orbitron has these pages already. Orbitron requires this package and publishes its kinetis://docs/* resources and both its tools from that one connection, so there is no second server to configure. Install and register kinetis/mcp-docs on its own when you want the documentation without the harness — beside a client, for any project or none.

Set up on its own, for Claude Code or Codex#

For Claude Code:

curl -fsSL https://raw.githubusercontent.com/kinetis-dev/kinetis/main/packages/mcp-docs/setup.sh | bash

For Codex:

curl -fsSL https://raw.githubusercontent.com/kinetis-dev/kinetis/main/packages/mcp-docs/setup.sh | bash -s codex

The script needs:

a running Docker daemon;
the claude or codex CLI on your PATH;
network access to Docker Hub, Packagist and GitHub.

It needs no PHP, Composer or sudo: the install and the server both run in the composer:2 image as your own user.

The script installs the package into ~/.kinetis-mcp-docs, checks that the server answers through the exact command it registers, and registers that command as kinetis-docs — at user scope in Claude Code. Start a new session to use it. Set KINETIS_MCP_DOCS_DIR to install elsewhere; the directory must be empty or an earlier install made by this script. Running the script again reinstalls in place.

Each session start checks for a newer 1.x release at most once a day, and a failed check still starts the installed server. Appendix: MCP Documentation Server Reference covers the directory marker, locking and the update lifecycle.

Running it directly#

Install the package with Composer, on PHP 8.4 or later:

composer require kinetis/mcp-docs
php vendor/bin/kinetis-mcp-docs

Register vendor/bin/kinetis-mcp-docs with any MCP client that launches a stdio server as a subprocess. It speaks MCP 2025-06-18, reads one JSON-RPC message per line on stdin, writes one response per line on stdout, and exits when its input closes. Diagnostics go to stderr.

Resources#

Each page is one resource at kinetis://docs/<slug>, where the slug is the page’s file name: kinetis://docs/routing-validation is Routing & Validation, and kinetis://docs/queue-sql is Queue (SQL). resources/list returns every page, and resources/read returns one:

printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"resources/read","params":{"uri":"kinetis://docs/routing-validation"}}' \
    | php vendor/bin/kinetis-mcp-docs

The response is one line whose result.contents[0].text is the page’s markdown, with mimeType text/markdown.

The list of pages is fixed in each release of the package; the server does not discover pages. A page added to the site after your installed release is not listed until the package is updated, and a URI outside the list is refused.

Reading a page in windows#

kinetis_read_doc returns one bounded window of a page, which is how a page is read — some pages run to a few thousand lines of markdown, and a window keeps a read to the part the task is about:

printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"kinetis_read_doc","arguments":{"uri":"kinetis://docs/appendix-routing-validation","startLine":1,"lineCount":200}}}' \
    | php vendor/bin/kinetis-mcp-docs

It takes the page uri, an optional startLine (from 1, default 1) and an optional lineCount (1 to 200, default 200). The result is one JSON document reporting status, uri, startLine, endLine, hasMore and content. Read on by calling again with startLine set to the reported endLine plus one, and only while what you came to the page for is still unresolved; concatenating the windows of one page reproduces it byte for byte, as long as the page has not changed on main between your calls.

A window ends at lineCount lines or 32 KiB of content, whichever comes first, so endLine can fall short of what you asked for — read it rather than assuming. hasMore and endLine describe the window you got. Every call fetches the page again, so they describe that call alone: nothing is cached, and no cursor or snapshot is held between calls.

Reading the same URI as a resource returns the page whole, for when the complete page is what you need.

Searching a page#

kinetis_search_doc finds the lines of one page that contain a literal string, so a named term in a long page is located with one call instead of window after window:

printf '%s\n' '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"kinetis_search_doc","arguments":{"uri":"kinetis://docs/appendix-database","query":"DB_REPORTING_DRIVER"}}}' \
    | php vendor/bin/kinetis-mcp-docs

It takes the page uri, a query of 1 to 256 characters, and an optional startLine (from 1, default 1). The result is one JSON document reporting status, uri, query, startLine, matches and hasMore; each match is a line number and that line’s content. Matching is literal and case-sensitive, and a call returns at most 50 matches: while hasMore is true, call again with startLine set to the last reported line plus one. Then read a window around the line you need with kinetis_read_doc.

The full argument, refusal and bound contract is in Appendix: MCP Documentation Server Reference.

How a page is read#

Every read — a whole resource, one window or one search — fetches the page from the main branch of the kinetis-dev/kinetis repository over HTTPS and works on its MyST markdown source as published — not rendered HTML, not a summary. There is no source, branch or path to configure, no local copy, and no cache, so a read returns what main holds at that moment. That can describe behavior newer than the Kinetis release your project pins.

A page that cannot be fetched answers a JSON-RPC error, and the URL and reason go to stderr. The fetch’s TLS, redirect, timeout and size limits are in Appendix: MCP Documentation Server Reference.