mcpTools

Returns the tools of the Model Context Protocol server that server is connected to, as tools of an agent. Many services publish such a server, so an agent gets their tools without a Tool written by hand for each.

// The app connects, with the Kotlin SDK of MCP. Over HTTP this works on every platform.
val server = HttpClient().mcpStreamableHttp("https://mcp.example.com/mcp")

val agent = toolAgent(model, tools = mcpTools(server) + ownTools)
agent.invoke(AgentState("Which issues are open?"))

server.close()

The function asks the server for all of its tools once, page by page. Each tool keeps the name, the description and the input schema the server gave it, and running it calls the server. Call this function before you build the graph, and close server when the app is done with the agent: the connection belongs to the app. A server that changes its tools later is not followed: call the function again and build the agent again.

What a tool returns to the model:

• The text parts of the result, joined. An image, a sound or a file becomes one line that says what it was, such as [image: image/png].

• A result the server marks as an error is an error result for the model, like a tool of the app that throws. So is a call that failed or got no answer in time. The run goes on, and the model can react.

To let an agent use only some tools, pick them with filter, which gets what the server says about each:

val readOnly = mcpTools(server) { tool -> tool.annotations?.readOnlyHint == true }

To approve the calls by hand, run the graph with interruptBefore = setOf("tools"), as for the tools of the app.

A model accepts letters, digits, _ and - in the name of a tool. Any other character of a name, such as the dot of repo.search, is given to the model as _, and the server is called with the real name.

Parameters

Throws