Ten thousand wires

About this video

Leila wants her agent to read her team's project tracker: what's due this week, and which of those are hers? Out of the box, the agent can't see the tracker. An agent uses tools: the harness tells the model which tools exist, the model writes a tool call, and the harness runs it. So someone has to write the code that talks to the tracker. If each agent's makers wrote their own, ten agents and a thousand services would need ten thousand connectors, and every change to a service would break ten of them.

MCP, the Model Context Protocol, is one agreed way to connect agents to tools and data. A service writes one MCP server, which offers tools (actions), resources (data to read) and prompts, and any agent whose harness speaks MCP, the client, can use it. Anthropic published MCP in November 2024; in December 2025 it was handed to a foundation run by the Linux Foundation, so no single company controls it.

A connection goes like this. The agent asks the server for its list of tools, with each tool's name, description and inputs. The model sees them like any other tools, and when it writes a call to one, the harness passes it to the server, which does the work and returns a result. The result goes into the context and the loop carries on. To the model nothing is new: MCP only changes who writes the tool.

A server runs on your computer, talking through standard input and output, or on the web. In Claude Code a team can share servers in a .mcp.json file in the project. But a server is someone else's code: a local one runs with your permissions, and whatever any server returns goes straight into the agent's context. Connect what you use, prefer the official server, and give it the least access it needs.

Transcript

1. 0:00Leila wants her agent to read her team's task tracker. Out of the box, it can't see it. Someone has to write the code that talks to the tracker. But who?
2. 0:22If every agent's makers write their own, ten agents and a thousand services need ten thousand connectors. And each time a service changes, they break.
3. 0:46MCP, the Model Context Protocol, is one agreed way to connect. Each service writes one MCP server, and every agent that speaks MCP can use it.
4. 1:10When the agent connects, it first asks the server which tools it has. The server lists them. The model calls one, the server does the work, and the result goes into the context.
5. 1:40To the model, nothing is new. It reads tool descriptions and writes tool calls, just as before. MCP only changes who writes the tool.
6. 2:00A server can run on your computer or on the web. A team can share its servers in a dot mcp dot json file in the project.
7. 2:20But a server is someone else's code, and whatever it returns goes straight into your agent's context. Connect what you use, prefer the official server, and give it the least access it needs.
8. 2:40One server per service, and every agent can plug in.