AI Developer Toolkit MCP Server

The AI Developer Toolkit MCP server is a public, read-only Model Context Protocol endpoint at https://developertoolkit.ai/mcp that exposes every guide on this site through two tools, search and fetch. No account and no API key are needed, and any client speaking Streamable HTTP can connect — Claude Code, Cursor, Codex and the ChatGPT connectors included.

Agents guess about tooling. They invent Cursor flags, misremember which file Claude Code reads its MCP config from, and confidently hand you a Codex option that was renamed two releases ago. The correction usually arrives after the command has already failed. Pointing the agent at a documentation server moves that check earlier: it reads the guide while it is still composing the answer.

Add the server

The endpoint is remote and hosted, so there is nothing to install and no process to keep alive.

Cursor

.cursor/mcp.json

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

Claude Code

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

Codex

~/.codex/config.toml

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

There is no authentication step afterwards. Most remote servers send you through an OAuth flow on first use; this one serves the same documentation to everybody, so it skips that.

The two tools

search takes a query and returns up to ten ranked results:

{ "results": [{ "id": "...", "title": "...", "url": "...", "snippet": "...", "lang": "en" }] }

fetch takes an id from those results and returns the whole article as markdown, plus metadata naming the language, the section, the last-updated date, and whether the page sits behind the subscription on the site.

The names matter more than they look. ChatGPT and Claude connectors expect exactly this pair, and each payload comes back both as structuredContent and as JSON inside the text content — which is what makes their citations resolve to a real URL instead of a paraphrase.

Tip

Copy-paste prompt once the server is connected:

Search the AI Developer Toolkit docs for how to make a Claude Code hook
block a commit, fetch the article, and then write the hook for this repo
following exactly what it says.

Ranking, and why Polish queries work

Scoring weights a title match five times a description match, then divides by how common the term is across the corpus. That second half is what separates the languages: an English query matches English titles, and a Polish query matches Polish ones, because the rare word carries the weight and rare words rarely cross a language boundary. There is no language parameter to set.

The EN and PL variants of the same article are collapsed into one result, so a query never spends half its ten slots on the same guide twice.

Limits worth knowing

• Read-only. Two tools, both reads. Nothing here writes to a repository, runs a command, or reaches beyond this site’s documentation.
• Stateless. Every call is independent and nothing is stored between them — which is also why there is no account.
• Not the only entry point. The same text is published at /llms.txt and /llms-full.txt. MCP exists because downloading six megabytes to answer one question is not reading.

Must-Have MCP Servers for Every Developer The five categories that anchor a working setup, and how to configure each one per tool.

MCP Best Practices: Optimization and Performance Keeping a stack of servers fast and context-efficient once more than one is connected.