bulaev.ai

Model Context Protocol

CCCOpen source

Agents · Personal & corporate AI OS

The open standard that lets any AI app connect to your tools and data; "USB-C for AI applications".

Open modelcontextprotocol.io

What it is

The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external systems. An MCP server exposes tools (functions the model can call), resources (data it can read) and prompts (reusable workflows); an MCP client such as Claude, ChatGPT, Claude Code, Cursor or VS Code connects to servers and lets the model use what they expose. The site's analogy is USB-C: build a server once and every client can plug into it. Transport is local (stdio) or remote (HTTP), with an authorization spec for remote servers, and official SDKs exist for the major languages; a newer "MCP Apps" extension lets a server ship interactive UI that renders inside the client.

In practice this is why Linear, Playwright, Publora, Apify, Notion, Context7 and Punctual in this catalog each have "an MCP server": it is the one integration that makes a tool usable by agents without bespoke code.

Where it fits in the AI OS

Personal & corporate AI OS, the wiring standard. The personal side: your assistant can see your calendar, notes and tasks. The corporate side: which servers are allowed, with which credentials, for which people; MCP is where access control and cost of an AI OS become concrete decisions rather than vibes.

Insider notes

members

How Serge's stack uses it

The stack is MCP-first on both ends. Publora ships an MCP server so agents can publish, and Punctual ships one so agents can book; on the consuming side, the project's Claude Code sessions use MCP servers for Linear, Playwright (browser inspection for the ideas pipeline), Context7 documentation and others, configured per project. Product rules in the spec follow the same idea: public entities are served as Markdown for agents, and /llms.txt lists the sections.

Setup in 5 minutes

  1. Pick one server you already need, for example the Playwright server: claude mcp add playwright -- npx @playwright/mcp@latest.
  2. Run claude, type /mcp to confirm it is connected and see its tools.
  3. Ask for something that needs the browser: "open example.com and tell me the h1".
  4. Add a remote server (Linear's hosted MCP URL) and complete its OAuth flow.
  5. Write your own: the TypeScript SDK's quickstart is one file that registers a tool() with a Zod schema and runs over stdio.

Alternatives

  • Function calling with hand-written tool schemas — per-app, no reuse across clients.
  • OpenAPI-based plugins / Actions — good for REST APIs, no local tools or resources.
  • Agent frameworks' own tool registries (LangChain, CrewAI) — work inside one framework only; most now also speak MCP.