M × N was the problem. M + N is the protocol.
10 apps × 10 tools, wired by hand 100 one-off connectors
with MCP, 10 + 10 20 connectors, total
The connectors do not disappear. Each one is written once, against the protocol, instead of once per pairing.
An MCP server exposes three things.
MCP, the Model Context Protocol, is an open standard for plugging tools and data into AI apps. Write a server once and Claude, ChatGPT, VS Code, and Cursor can all use it.
Functions the model can call.
search_issues(query)
Data the model can read, addressed by URI.
file:///repo/README.md
Prompt templates the user can invoke.
/review-pr
Transports: stdio for local servers, Streamable HTTP for remote.
That’s the whiteboard. The one-pager is the handout.
Email me the one-pagertotal: 0.00. unsubscribe: one click.
move the pointer across this panel: the glue code is still underneath