Model Context Protocol

MCP is an open protocol, introduced by Anthropic in late 2024, that standardises how AI applications connect to external systems. A client — an agent, an editor, a desktop app — speaks JSON-RPC to an MCP server, which advertises what it can do.

Before it, every integration was bespoke: one adapter per agent per service. MCP makes the service implement one interface and every compliant client gets it.

The three primitives

Primitive Who decides to use it Shape
Tools The model Named function + JSON input schema
Resources The client Readable data pulled in as context
Prompts The user A reusable template, often a slash command

The distinction matters for design. A tool is something the agent may call on its own initiative, so its schema and description are effectively prompt text. A resource is something the client fetches deliberately.

Transports

  • stdio — the client launches the server as a subprocess and talks over standard input and output. Local, no network, no auth beyond process boundaries.
  • Streamable HTTP — a remote server over HTTP, with streaming responses. This is what a hosted service exposes.

Same protocol, different framing.

Authorisation is the server's business

MCP does not invent an auth model. A remote server applies whatever it already applies: a bearer token per user, with the same expiry, revocation, scope and rate-limit rules as its REST API.

That is the right shape for agents specifically, because an agent is software you are handing a credential to. A token that is scoped to publishing, revocable in one click, and rate-limited per token is a very different risk from a session cookie (scoped tokens for AI agents).

What this changes for reports

An agent that finishes a long task has output. Pasting it back into the chat is the default, and it is the wrong shape: chat has a message ceiling, no formatting beyond markdown, no place for anyone else to respond, and it scrolls away (the agent's output is too long to paste).

With an MCP server for publishing, the agent creates a report as a normal step in its run and returns a URL. The artifact outlives the session, has an address a teammate can open, and can be commented on — and the agent can read those comments back on the next run and act on them (let an agent respond to comments).

Comma's MCP server wraps the same surface as the REST API, authenticated with the same comma_sk_… tokens, so scope, revocation and rate limits behave identically either way.

Try it

Comma is free — unlimited reports, unlimited commenters, unlimited revision history.

Set up the MCP server →

Related