Using MCP Tools in Your Own Agent

The client's job

When you build your own host, your MCP client does three things: connect to the server, call tools/list to get the tool definitions, and forward each model-requested call with tools/call. The definitions slot straight into the tool list you already send to the model, so the agent loop from Section 1 does not change.

The idea in pseudocode

Real SDK calls differ by language and version, so this sketch shows the flow rather than exact method names.

session = connect(server_command)           # 1. connect
tools = session.list_tools()                # 2. discover
model_tools = [to_model_schema(t) for t in tools]

# inside the agent loop, when the model requests a tool:
result = session.call_tool(name, arguments) # 3. forward the call

Namespace tool names

Two servers may both expose a tool called search. Prefix names with the server (for example issues_search) when merging tool lists so the model and your router never confuse them.

Quick check: Which MCP method returns the available tool definitions?

• tools/call
• tools/list
• prompts/run
• server/delete