Notes

What is MCP (Model Context Protocol)?

Updated April 2026

MCP — short for Model Context Protocol — is an open standard from Anthropic that lets AI assistants like Claude connect to tools and data sources through one common interface. MCP servers can run locally on your machine over stdio, or remotely as hosted services over HTTP. Instead of building a custom integration for every app, a server exposes its capabilities once and any MCP-compatible assistant can use them.

What it is

MCP, the Model Context Protocol, is an open standard published by Anthropic in November 2024. It defines a JSON-RPC interface that AI assistants (clients) use to talk to outside systems (servers) — your filesystem, a database, a meeting library, a calendar — without each side knowing anything specific about the other.

A useful analogy: MCP is to AI assistants what USB-C is to laptops. Before USB-C, every device had its own connector. After it, one cable works everywhere. MCP plays the same role for tool integrations.

Why it matters for end users

If you’ve ever wanted Claude or ChatGPT to do something with a specific app — read your meeting notes, query a private database, search your codebase — you needed someone to build a custom integration. With MCP, that work happens once, server-side, and every MCP-compatible chat client gets it for free.

Concretely, that means:

  • You set up the MCP server for the tool you care about — for hosted servers (like Granola’s), through a settings or OAuth flow inside the chat client; for local servers (like Gistlist’s), by adding an entry to the client’s config file (e.g. claude_desktop_config.json).
  • You connect your chat client to it.
  • The assistant can now answer questions and take actions that involve that tool, with no copy-paste.

It’s the same shift that browser extensions enabled for the web: the assistant becomes useful for your data, not just generic web knowledge.

Examples of MCP-enabled tools

Anthropic’s reference server collection and community catalogs together cover a wide range of integrations. A few public examples:

  • Filesystem — let Claude read and write files in a folder you choose.
  • GitHub — query repos, issues, and pull requests through Claude.
  • Slack — search and post to your workspace.
  • Postgres / SQLite — query a database in natural language.
  • Gistlist — search across your local meeting library and pull the relevant transcripts into the conversation. (See /notes/founder-one-on-ones for a worked example.)

Many MCP servers — including Gistlist’s — run as local subprocesses on your machine, communicating over stdio. Others, like Granola’s, are hosted on the vendor’s infrastructure and reached via remote HTTP. Both patterns are part of the spec; which one a given server uses determines where your data flows.

How MCP differs from plugins, extensions, and APIs

It’s easy to mistake MCP for the latest in a long line of integration mechanisms. The substantive difference is where the contract lives.

MechanismContract is between…Reusable across clients?
Custom API integrationThe app and one specific clientNo
Browser extensionThe browser and one website/featureNo (browser-locked)
ChatGPT plugin (deprecated)OpenAI and the plugin authorNo (OpenAI-only)
MCP serverAny MCP client and any MCP serverYes

The reusability is the point. A single Gistlist MCP server works with Claude Desktop today and with any future client that adopts the spec — no extra integration work.

Using MCP with Gistlist

Gistlist ships an MCP server that exposes three tools to Claude Desktop: list_recent_meetings, search_meetings, and get_meeting. Claude can use them to answer questions across your whole meeting history — “what did we decide in the Acme onboarding retro?”, “who have I met with this month?” — without you copying transcripts into the chat. Setup is a config-file edit; details are in the Claude Desktop integration docs.

The fit is natural: Gistlist already keeps every meeting as plain markdown on your disk, so there’s no separate import or sync step to a Gistlist-hosted service. The local MCP server reads the on-disk index and returns the meetings Claude asks for; that selected content does become part of Claude’s conversation context, since that’s how Claude answers your question.

Further reading

If you’re comparing meeting tools that integrate with Claude, Granola vs Gistlist walks through how the two products differ — including how each one handles MCP.

If you want a concrete workflow, Tracking founder 1:1s without a CRM shows what asking Claude across an MCP-connected meeting library actually looks like in practice.

Frequently asked questions

Is MCP a Claude-only thing?

No. Anthropic published MCP as an open specification in late 2024, and clients beyond Claude Desktop have been adopting it. The protocol itself is not tied to any one model or vendor — any chat client can implement it, and any tool or data source can expose itself through an MCP server.

Do I need to code to use MCP?

No. Using an existing MCP server is a setup step inside the client you're using — for hosted servers like Granola's it's a settings/OAuth flow, and for local servers like Gistlist's it's a small JSON entry in the client's config file (e.g. claude_desktop_config.json for Claude Desktop). Building a new MCP server requires code; using existing ones does not.

Is MCP secure?

It depends on the server, the transport, and what permissions you grant. Many desktop MCP servers (Gistlist's included) run as local stdio subprocesses launched by a client you trust, so the data path is local. MCP also supports remote HTTP transports for hosted servers — those send your prompts and the server's responses over the network like any other API. Security is a property of the specific client + server + transport you've configured, not of the protocol itself.

Does ChatGPT support MCP?

OpenAI added MCP support to ChatGPT in 2025 under the "connectors" feature. Other clients including Cline, Zed, and a growing list of IDE assistants also speak MCP. The protocol is becoming the common adapter for tool-augmented chat.

What's the difference between MCP and a regular API?

An API is a contract between two specific systems. MCP is a contract between any model and any tool — the same way USB-C lets any laptop talk to any monitor without custom cables. An MCP server can be plugged into any MCP client, no per-client integration work required.

Where can I see the official spec?

The full specification, reference implementations, and a list of community-built servers are at modelcontextprotocol.io.

← All notes