mcpField
R
runitonai-mcp

runitonai-mcp

Remote
runitonai.comMemory & Knowledge0Updated 13w ago

About runitonai-mcp

An open-source reading companion server for the Run It on AI book, intended for readers connecting an AI client to its chapters and exercises. Its source is published for inspection; Implementation Blocks are available without a license key, while chapter text requires one.

Key features

  1. 01Serves book operating instructions
  2. 02Provides licensed chapter text
  3. 03Serves Implementation Blocks without credentials
  4. 04Published server source code
  5. 05Rate-limits license keys

Use cases

  1. 01Retrieve book chapters through an AI client
  2. 02Access book exercises without a license key
  3. 03Inspect the companion server's code

MCP support

Provides MCP tools through https://mcp.runitonai.com/mcp for book content. It can be connected from Claude, Claude Code, and Codex; ChatGPT can use its OpenAPI description as an Action.

Configuration

Remote servers need no local install. Pick your client, then copy the config or install in one click.

Endpointhttps://mcp.runitonai.com/mcp
Copy install command
Claude CodeTerminal
Cursor.cursor/mcp.json
VS Code.vscode/mcp.json
Claude Desktopclaude_desktop_config.json
Windsurf~/.codeium/windsurf/mcp_config.json

Run commands in a terminal; paste JSON into the client’s MCP config and restart. Servers that need sign-in will prompt on first connect.

Tools

11

This tool takes no parameters.

Details

From the project README

Run It on AI — the companion server, in full

This repository is the complete source code of the server that powers the Run It on AI reading companion at mcp.runitonai.com. If you bought the book and connected your AI to it, this is everything that runs on our side.

We publish it so you do not have to take our word for anything. The server is a few hundred lines of TypeScript. You (or your AI — paste any file into it and ask) can read exactly what it does and, just as importantly, what it cannot do.

The short version

What reaches our server when you use the companion:

  • The name of the tool your AI called (for example "fetch chapter 3").
  • The chapter or exercise it asked for.
  • Your access code — the license key printed on your Gumroad receipt.

What never reaches us, because there is no code here that could receive, store, or forward it:

  • Your conversations with your AI.
  • Your files, documents, or anything on your machine.
  • Your name.
  • Your email. Search this repository for an email field: there is none. No tool, no endpoint, no code path accepts one.
  • The picture your AI forms of you while you work. That stays inside your AI, on your side.

Who talks to whom: when you add the companion as a Claude connector, your machine never contacts our server directly — Anthropic's cloud brokers the connection. On ChatGPT, OpenAI's cloud does. We see tool calls, not you.

How access works

Your credential is the license key on your Gumroad receipt (a code like XXXXXXXX-XXXXXXXX-XXXXXXXX-XXXXXXXX). It identifies a purchase, not a person. The server checks it against Gumroad, caches a hash of it for a day, and rate-limits it. Refunded or disabled keys stop working. See src/license.ts — the whole mechanism is one file.

Only chapter text needs the key. The companion's operating instructions (activate) and every Implementation Block — the book's working exercises — are served without any credential at all; the exercises are also published right here in implementation-blocks/.

Connect it

Claude (web or desktop) — the easy path. Settings -> Connectors -> Add custom connector, then paste:

https://mcp.runitonai.com/mcp

Start a chat with: Activate Run It on AI — my access code is <your license key>.

ChatGPT. Use the "Run It on AI Companion" GPT (link in the book), or build your own: create a Custom GPT and import https://mcp.runitonai.com/openapi.yaml as an Action.

Claude Code (terminal users).

claude mcp add --transport http runitonai-book https://mcp.runitonai.com/mcp

Codex. In ~/.codex/config.toml:

[mcp_servers.runitonai_book]
url = "https://mcp.runitonai.com/mcp"

What is in this repo

PathWhat it is
src/worker.tsThe Cloudflare Worker: routing, transports, the Gumroad sync endpoint
src/license.tsLicense-key verification and caching (the access gate)
src/auth.tsVerification of legacy bearer tokens (nothing mints new ones)
src/rate_limit.tsPer-key rate limiting
src/mcp/MCP protocol plumbing (tool catalog, dispatch, transports)
src/tools/The tools your AI can call
openapi-v1.yamlThe same surface described for ChatGPT Actions
implementation-blocks/Every Implementation Block from the book, as readable markdown

Two data files ship as placeholders: src/data/chapters.json (the book's chapter text) and src/data/skill.json (the companion's operating instructions). Both are generated from the book manuscript at deploy time. The book's prose is the product, so it is not in the public repo — but the code that serves it is all here, and the Implementation Blocks (the working exercises) are published in full in implementation-blocks/.

Note for contributors: everything in this repo is generated from the private manuscript repo by a sync script, and a pre-commit hook there re-syncs it on every relevant change. The Implementation Blocks live in the book's manuscript; the copies here are build artifacts — never edit them here, they will be overwritten.

Run it yourself

npm install
npm run typecheck
npm run dev        # local server on localhost:8787

wrangler.toml ships with placeholder KV namespace ids; point them at your own Cloudflare account if you want to self-host the shell.

License

MIT — see LICENSE. The book itself (chapters, companion instructions) remains copyrighted content of StructLabs.io Pte Ltd and is not part of this repository.

Comments

Markdown supported

New to MCP?

From what the protocol is to installing your first server, these answers cover most first questions.

An open protocol that defines how AI applications connect to external tools and data in a uniform way. Clients talk to the model; MCP servers provide tools, resources and prompt templates. They exchange standard messages, so one server can be reused by many clients.