MCP in one paragraph
The Model Context Protocol (open-sourced by Anthropic in Nov 2024, since adopted across major AI platforms) is a standard way for AI applications to plug into external systems. A host (the AI app) runs clients that connect to servers; each server publishes tools (actions), resources (data), and prompts. It's often called "USB-C for AI": build a server once, and any MCP-capable agent can use it — no bespoke integration per model.
Tools
Callable actions with typed schemas: search_posts, reply, create_offer. The agent reads the schema and decides when to call.
Resources
Readable context — timelines, order history, ratings — that the host can pull into the model's context window on demand.
Hosted vs local
Local servers run on your machine (files, git). Hosted servers run at the platform (X, a marketplace), so agents anywhere can connect over HTTP with OAuth-style auth.
The agent-commerce loop
With Social + Marketplace + Payments servers connected, a service agent can close a full deal:
1. Discover — search_posts("need a landing page") finds live demand.
2. Qualify & pitch — pull verified ratings and completed-order counts from the marketplace server, reply with proof and a quote.
3. Transact — attach an x402 payment link; buyer's agent pays; work is delivered and rated.
Worked numbers: if an agent scans 5,000 relevant posts/day, pitches the best 2% (100), and converts 5%, that's ~5 jobs/day. At a $40 average order it clears $6,000/month — with the marginal pitch costing about a cent of tokens. That unit economics is why "agent commerce" is suddenly a serious category.
What to watch (and watch out for)
Standards fights (MCP vs proprietary plugin APIs), spam pressure — the same loop that finds demand can flood it, so platforms rate-limit and require verified track records — and security: a malicious server's tool descriptions are prompt-injection surface, which is why hosts sandbox tool calls and require user approval for consequential actions.