MCP: How an AI gets access to its tools

You ask an AI about the METR post on this website. The model can compose an explanation. But the current post has to enter its context first. The application needs a concrete capability: searching posts.

The search receives METR and English. It compares them with published titles, summaries and sections. It returns matching posts and their identifiers. The next tool receives one of those identifiers, reads the post and returns its sections. Now the model can base its explanation on that text, with a source you can check.

Model Context Protocol, or MCP, is a common connection language between an AI application and a service. The application contains an MCP client; the service provides an MCP server. Here, the service is this website. MCP describes how the application discovers capabilities, sends an invocation and receives its result. MCP architecture.

A tool has a name, a description and an input format. Through tools/list, the client discovers the available tools. Through tools/call, it asks the server to execute a named tool with appropriate arguments. The server executes the function and returns a result or an error. Tools in the protocol.

The chat behind “Ask AI” calls our tools directly through the model API. It does not use MCP for these calls. Separately, we expose the same three reading functions to other AI applications at https://appliedbyhand.com/mcp:

• list_posts takes optional filters and returns a list of posts.
• search_posts takes search terms and returns matching posts and sections.
• get_post takes a post identifier and returns its full text.

The public connection provides published content. Browser actions such as changing appearance, preparing a contact draft or opening a self-test belong to the website chat. For contact and newsletter forms, you review and submit yourself. When a sandbox is configured, the chat can also execute code examples separately.

Open “Ask AI” → “This AI’s toolbox” above to inspect the capabilities actually offered.

MCP does not create the language model or the search function. It provides access to existing functions. The application and service still determine whether those functions produce useful results, which data they can access, and which actions you allow.

In the demo above, you run the reading calls yourself. You see real results from the connection; a model would then compose the explanation. If the search finds no post, there is no source text from that result to support a claim.

If your application supports a remote MCP server through Streamable HTTP, add https://appliedbyhand.com/mcp as its server URL. Our connection negotiates protocol versions 2025-06-18 and 2025-03-26. Your client must support one of them. No API key is needed for the published reading tools. The available tool list defines what you can access. Initialization, HTTP transport.