smolagents

Install

pip install "smolagents[mcp]" "mcp<2"

The server runs in its own environment through uvx, which needs uv. The mcp<2 pin applies only to your agent's environment. smolagents[mcp] 1.26.0 installs mcpadapt 0.1.20, which fails to import under mcp 2.x with ImportError: cannot import name 'streamablehttp_client'. The tradefloor server still gets the mcp 2.x it needs, because uvx installs it separately.

Connect

MCPClient starts the server as a subprocess, returns its tools as a list and closes the connection when the with block exits.

import json

from mcp import StdioServerParameters
from smolagents import MCPClient

server = StdioServerParameters(
    command="uvx",
    args=["--with", "tradefloor[mcp]", "tradefloor", "mcp"],
)

with MCPClient(server, structured_output=False) as tools:
    names = sorted(tool.name for tool in tools)
    print(len(names), "tools:", ", ".join(names[:4]), "...")

    list_scenarios = next(tool for tool in tools if tool.name == "list_scenarios")
    result = json.loads(list_scenarios())
    print("scenarios:", ", ".join(s["name"] for s in result["shipped"]))

13 tools: build_scenario, build_universe, check_envelope, check_job ...
scenarios: curve_shock, geopolitical_conflict, liquidity_crisis, oil_price_spike, policy_regime_shift, rate_shock, recession

Setting structured_output explicitly silences a FutureWarning about its default changing. With False, each tool returns the server's JSON as a string.

To give the tools to an agent, pass the list inside the with block, for example CodeAgent(tools=tools, model=model) or ToolCallingAgent(tools=tools, model=model). Both agents also add their own final_answer tool. ToolCollection.from_mcp(server, trust_remote_code=True) loads the same 13 tools, and it raises a ValueError unless you pass trust_remote_code=True to confirm you trust a server that runs code on your machine.

Tested with

smolagents 1.26.0, mcpadapt 0.1.20, mcp 1.30.0, Python 3.12.14, uv 0.12.14 and tradefloor 0.9.1.