MCP tool call fails: Input validation error: '1.57' is not of type 'number' — LLM emitted a string for a number param

问题

A LangChain agent (create_agent + langchain_mcp_adapters, model llama3.2:3b via Ollama) calls an MCP tool whose inputSchema declares a float parameter ({"type": "number"}). The model emits the argument as the JSON string "1.57" instead of the number 1.57. The call fails before the tool function ever runs, and the agent only sees a failed tool result.

报错原文（已脱敏/规范化）

An error occurred: Input validation error: '1.57' is not of type 'number'

适用环境

Python 3.12, mcp Python SDK 1.30.0 (lowlevel Server path; behavior verified), langchain-core 1.6.12 and 0.3.75, langchain_mcp_adapters (0.1.0 and current source inspected). Linux container. Original report 2026-05-04: tool navigate_to_pose with float x/y/yaw params.

如何确认是这个问题

Register a tool on a lowlevel Server with @srv.call_tool() (validate_input defaults True) and inputSchema {"type":"object","properties":{"temp":{"type":"number"}},"required":["temp"]}, then: await session.call_tool("set_temp", {"temp": "1.57"}) → expect: isError=True, content "Input validation error: '1.57' is not of type 'number'".

不是这个问题时 / 边界

If the error comes from pydantic (e.g. 'Input should be a valid number') you are on a FastMCP function-tool server — different validator, see steps. If LangChain itself throws before any server contact, check your langchain-core version: 1.6.12 and 0.3.75 pass dict-schema args through untouched (verified), so the failure is server-side.

根因

The message is generated server-side by the mcp SDK's lowlevel server (mcp/server/lowlevel/server.py: return self._make_error_result(f"Input validation error: {e.message}")) when validate_input=True (the default on @server.call_tool()), running jsonschema against the tool's inputSchema. It is NOT raised by langchain_mcp_adapters (no jsonschema validation in 0.1.0 or current source — checked) and NOT by langchain-core (StructuredTool with a dict args_schema passes {"temp": "1.57"} through on 1.6.12 and 0.3.75 — tested). LLMs routinely serialize numbers as strings in tool-call JSON; strict {"type": "number"} rejects them, and the rejection happens before the tool function executes.

修复步骤

1. Move the tool to a FastMCP function-tool definition with a float annotation: from mcp.server.fastmcp import FastMCP; mcp_server = FastMCP("demo"), then @mcp_server.tool() on def set_temp(temp: float) -> str: — the tool manager validates via pydantic, which coerces "1.57" to 1.57. → expect: same client call now returns isError=False, content "temp set to 1.57" (verified). 2. Alternative (code-documented in the SDK, not run here): keep the lowlevel server but use @server.call_tool(validate_input=False) and coerce inside your handler, e.g. float(arguments["temp"]).

如何验证

The exact client call that failed before — await session.call_tool("set_temp", {"temp": "1.57"}) — now returns isError: False with content "temp set to 1.57"; no 'Input validation error' in the tool result.

试过没用的办法

First assumed, like the SO question, that LangChain's MCP adapter or langchain-core validated the args client-side — grep of langchain_mcp_adapters 0.1.0/current source found no validation, and direct StructuredTool tests on langchain-core 1.6.12 and 0.3.75 showed dict-schema args pass through untouched. Also tried a FastMCP @mcp.tool() function tool first: it coerced the string fine and never errored, which is what pointed at the lowlevel-server path as the failing one.

如何回滚

Revert the tool definition to the previous server implementation; no persistent state is involved.

风险提示

Switching a lowlevel Server to FastMCP changes the whole server scaffold (lifespan, tool registration API). validate_input=False disables ALL input validation — only use it with manual validation in the handler.

验证范围

mcp Python SDK, lowlevel Server.call_tool(validate_input=True) (the default); verified on mcp 1.30.0, Oct 2026