promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_promptdojo_

Wire an MCP tool — load tools from a server, not a registry — step 9 of 9

ch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path buildch 25 · mid-path build9/9
Checkpoint

One last thing before we move on. pass this to mark the lesson done, or skip and keep moving. hop to the next when you're ready.

Final drill of the capstone. Build the agent loop running with MCP-sourced tools. Write run_mcp_agent(question, tools_list_response, mcp_call, fake_model, max_iters) that:

  • Bridges tools_list_response via bridge_mcp_tools(...) (provided) into tools, schemas.
  • Loops up to max_iters times calling fake_model(messages, list(tools.keys())).
  • The model returns a dict like {"stop_reason": "...", "content": [...]} where each content block is either {"type": "text", "text": "..."} or {"type": "tool_use", "id": "...", "name": "...", "input": {...}}.
  • On end_turn: collect text blocks, return {"ok": True, "answer": <joined text>, "iters": <iter num>, "tool_calls": <count of tool_use blocks across all turns>, "tool_errors": <count of TOOL_ERROR results>}.
  • On tool_use: dispatch each tool_use block through tools[block["name"]](**block["input"]). Count any result that starts with "TOOL_ERROR:" as a tool error. Append assistant
    • user turns the standard way.
  • On cap: return {"ok": False, "error": "capped", "iters": max_iters}.

Two cases run for you. Expected output:

ok=True iters=2 tool_calls=1 tool_errors=0 answer=Found Tokyo's best ramen.
ok=True iters=2 tool_calls=1 tool_errors=1 answer=I couldn't search; tool was down.

full-screen editor opens — close anytime to keep reading.

open the editor to run it.