Your agent can read the guide before it guesses.

Your agent already guesses at Cursor flags, Claude Code hooks and Codex config. Point it at this server and it reads the matching guide in the middle of a task, without you leaving the editor.

  • Free
  • No account
  • No API key
  • Read-only

See a session

It runs on our server. Add it once and both tools appear in your next session.

Your tool

Install in Claude Code

Terminal

claude mcp add --transport http --scope user dtk https://developertoolkit.ai/mcp

Then check it. claude mcp get dtk health-checks the server, and /mcp inside a session shows each server’s status and tools.

Install in Cursor

.cursor/mcp.json

{
  "mcpServers": {
    "dtk": {
      "url": "https://developertoolkit.ai/mcp"
    }
  }
}

Then check it. Open the MCP section of Cursor’s settings and look for dtk, listed with its tools search and fetch.

Install in Codex

Terminal

codex mcp add dtk --url https://developertoolkit.ai/mcp

Or write it by hand: ~/.codex/config.toml

[mcp_servers.dtk]
url = "https://developertoolkit.ai/mcp"

The command writes this table into the file. Copy takes the command above, not the file.

Then check it. Run codex mcp list in a terminal and dtk appears with its URL. Inside a session, /mcp lists the tools the server advertises.

Each snippet is copied from the setup guide for this server.

ENDPOINThttps://developertoolkit.ai/mcp

1100+

articles, English and Polish

2

read-only tools, search and fetch

The index is rebuilt on every deploy, so answers match the live guides.

A question goes in. A cited guide comes back.

One session, step by step: what you type, the two tool calls your agent makes, and the answer it writes from the article.

  1. You ask

    A question in your editor, typed the way you would ask a colleague.

    My PreToolUse hook exits 1 when it sees rm -rf, but the command still runs. Why?

  2. The agent searches

    One call to search with a few keywords. It gets back titles, ids and short descriptions.

    search{ "query": "hooks exit code" }

    1. Claude Code hooks: guardrails, gates, and live context en/claude-code/automate/hooks-automation

      Claude Code hooks that block risky commands, protect paths, lint every edit, and refuse to stop on red tests, with fixture tests that prove each one works.

    2. Shared hooks governance: reviewed runtime controls en/cto-scorecard-guide/shared-hooks-governance

      Govern shared agent hooks like production code: a contract, pinned versions, fixtures, managed distribution, staged rollout, audit logs and a tested rollback.

    3. Use hooks as deterministic guardrails en/developer-scorecard-guide/agent-hooks

      Four layered hooks for Claude Code and Codex: protect paths, block risky commands, lint each edit, and refuse to stop with secrets or failing tests.

    Three of the ranked results shown

  3. It fetches the guide

    One call to fetch with the id from the first result. The whole article comes back as markdown.

    fetch{ "id": "en/claude-code/automate/hooks-automation" }

    title
    Claude Code hooks: guardrails, gates, and live context
    url
    https://developertoolkit.ai/en/claude-code/automate/hooks-automation/
    metadata
    lang en, section claude-code, lastUpdated 2026-10-12, gated true
    text
    The complete article as markdown. The passage the agent needs:
    Any other exit code, including 1, does not block unless valid JSON on stdout makes the decision. Claude Code shows a “hook error” notice and the action proceeds.
  4. It answers and cites

    The answer comes from the article, with its title and the date it was last updated.

    Exit 1 does not block a PreToolUse hook. Exit 2 does, and the text your script writes to stderr goes back to Claude as the reason. With exit 1 you get a “hook error” notice and the command still runs.

    Change the exit code to 2, or answer in JSON with a permissionDecision of deny inside hookSpecificOutput.

    Source: Claude Code hooks: guardrails, gates, and live context updated 2026-10-12

An illustration of one session. The tool names and arguments are the real contract. The three results, the guide, its id and its date are real articles, read from the library when this page is built. The question and the answer are written for this page, and the fetch result is shortened.

Two tools. That is the whole contract.

The names are the pair that ChatGPT and Claude connectors look for, so citations resolve on their own. Each search result carries the id you pass straight to the second tool.

search

in
{ query: string }
out
{ results: [{ id, title, url, snippet, lang }] }

Up to 10 ranked results across both languages at once. The English and Polish versions of one article count as one result. A query made of Polish words returns Polish articles, with no language parameter.

Each snippet is the article’s short description. Queries run to 1024 characters.

fetch

in
{ id: string }
out
{ id, title, text, url, metadata }

The whole article as markdown. metadata carries the language, the section, the last-updated date and a gated flag, which the FAQ below explains.

An unknown id is an error. Call search first.

How search ranks. It matches the words in a guide’s title, headings and description. A title match counts most, and rare words count more than common ones. It does not search article bodies, so ask with the terms the docs use.

Try search from your browser

This box calls the same endpoint your editor would. The results below are the real response: no mock, no cached sample.

Try

What it reads, and what it does not do.

Every guide on this site, in both languages, rebuilt on each deploy: Cursor, Claude Code and OpenAI Codex, from first setup through hooks, MCP, testing, CI and deployment.

It only reads

Two tools, both read-only. Nothing here can write to your repository, run a command, or reach anything but this site’s own documentation.

It keeps no session

Every call stands alone, so the server needs no account and no key.

Search is keyword matching

It returns at most 10 results and does not understand a question. If the first query misses, your agent can try the guide’s own terms, such as hooks or worktrees.

It is not the only way in

The same text is published at /llms.txt and /llms-full.txt for crawlers. MCP exists because downloading the whole corpus to answer one question is not reading.

The server is free. The Setup Pack and the books come with the Toolkit.

The Setup Pack builds your own .mcp.json from a catalogue of 31 MCP servers, each checked on 2 October 2026, so your agent connects to the servers that fit your stack. The Toolkit includes it, with all eight books and the full library for 12 months.

Eight books, at a comparable ebook’s price$383

The Toolkit, paid once$59.99

Books: eight titles at $47.99 each, the median list price of ten comparable ebooks on AI-assisted coding, with the total rounded down. The four role paths are made from the other books’ chapters in another order. The guide library is not counted.

$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.

Open the Setup Pack

It is included in your access, with all eight books and the full library.

Questions before you connect it.

Something else? Write to support@developertoolkit.ai.

Does it cost anything?

No. The server is free and needs no account or API key. The Toolkit adds the Setup Pack, the eight books and every guide in your browser.

Does fetch return guides that need the Toolkit on the site?

Yes. It returns the full text of every guide, including the ones that need the Toolkit on the site, and metadata.gated tells your agent which ones those are. The Toolkit adds the Setup Pack, the eight books and every guide in your browser.

Is it safe to connect to my agent?

Both tools are read-only and the server reaches only this site’s documentation. It cannot write to your repository or run a command, and it keeps no session between calls. The text it returns lands in your agent’s context.

Which tools and clients work with it?

Claude Code, Cursor and Codex, with the snippets above. Any MCP client that speaks Streamable HTTP works too, including ChatGPT connectors, where you paste the endpoint URL directly.

What do you record when my agent calls it?

Each call is recorded in our analytics (PostHog, EU region) under an anonymous ID: the tool, the query or id your agent sent, how long it took, the client’s user agent, and a location that PostHog derives from the caller’s IP address. An MCP client does not log in. A search from the box on this page is marked as coming from the page and, if you accepted analytics cookies, joins your analytics profile, so when you are signed in it is tied to your account. Keep secrets out of queries. The privacy policy covers the analytics.

Why did search miss what I asked?

Search matches the words in a guide’s title, headings and description. It does not read article bodies and does not interpret a question. Phrase the query with the terms the docs use, and let the agent try two or three phrasings.

Does it work in Polish?

Yes. Ranking covers both languages at once, and a query made of Polish words returns Polish articles with no language parameter. A query made only of English terms, such as Claude Code hooks, can return the English version, because English wins ties: add a Polish word, or ask the agent to fetch the pl/ id.

Let your agent look it up.

One URL, two read-only tools, and every guide in the library within reach of the next prompt.

The server is free. Toolkit: $59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.