[BUG] Tools parameters/arguements are not working

[AI3LabsX]

Bug Description

When
I use const tools = await client.getTools();
and
const bindLLM = llm.bind({
tools,
tool_choice: "auto",
});
it does not see tool arguments and always provide with empty arguments:
"tool_calls": [
{
"id": "call_Njw1orEyCrlaWWSjg3KTvdOE",
"type": "function",
"function": {
"name": "mcp__firecrawl__firecrawl_scrape",
"arguments": "{}"
}
}
]

Environment

• OS:Win

[AI3LabsX]

Code:

import { MultiServerMCPClient } from "@langchain/mcp-adapters";
import { ChatOpenAI } from "@langchain/openai";
import dotenv from "dotenv";
import util from "util";

dotenv.config();

(async () => {
    // Initialize the MCP client with server configuration for firecrawl.
    const client = new MultiServerMCPClient({
        // Global tool configuration options.
        throwOnLoadError: true,
        prefixToolNameWithServerName: true,
        additionalToolNamePrefix: "mcp",
        // MCP server configuration.
        mcpServers: {
            // Use a STDIO connection for the firecrawl server.
            firecrawl: {
                transport: "stdio",
                command: "npx",
                args: ["-y", "firecrawl-mcp"],
                env: {
                    FIRECRAWL_API_KEY: process.env.FIRECRAWL_API_KEY!,
                },
                // You can add additional options like timeout or restart configurations if needed.
            },
        },
    });

    try {
        // Load the tools from the Firecrawl MCP server.
        const tools = await client.getTools();
        // console.log("Loaded tools:", tools);
        console.log(util.inspect(tools[0], { showHidden: true, depth: null, colors: true }));

        // const result = await tools[0].invoke({ url: "https://github.com/mendableai/firecrawl-mcp-server", formats: ["markdown"], onlyMainContent: true });
        // console.log("Result:", result);
        const llm = new ChatOpenAI({
            modelName: "gpt-4o-mini",
        });

        const bindLLM = llm.bind({
            tools: tools,
            tool_choice: "auto",
        });
        const result = await bindLLM.invoke([
            // { role: "system", content: `Tool Prompt: ${JSON.stringify(tooldata)}` },
            {
                role: "system",
                content: `Use properties to get information about tool required arguments!`
            },
            {
                role: "user",
                content: "What is that page about: url: https://github.com/mendableai/firecrawl-mcp-server"
            }]);
        console.log("Result:", JSON.stringify(result.tool_calls, null, 2));

    } catch (error: any) {
        console.error("Error during agent invocation:", error);
        // For tool-specific errors, you might check error.name === "ToolException"
        if (error.name === "ToolException") {
            console.error("Tool execution failed:", error.message);
        }
    } finally {
        // Close the MCP client to clean up resources.
        await client.close();
    }
})();

results:
Result: [
{
"name": "mcp__firecrawl__firecrawl_scrape",
"args": {},
"type": "tool_call",
"id": "call_u0g1SHlRwjIdKz6bcdS5So3t"
}
]

In loog I can see properties are there, but bind can not pick them up.

[AI3LabsX]

Just to mention, the same behaviour is with CreateReactAgent.

[postbird]

Same issue, I think the issue is caued by DynamicStructuredTool.

And I tried directly use DynamicStructuredTool to build the ReActAgent just like what this repo implemented, for example:

And I pass this tool to ReAct agent like below:

And I found that the LLM can not process the tool params, when I debug with the langsmith, I find that actually the params of the function calling are invalid.

[AI3LabsX]

So I fixed my issues by down grading version a bit. Also am binding the whole getAllTools (maybe with some prefilteting step) and it works nicely.

[postbird]

So I fixed my issues by down grading version a bit. Also am binding the whole getAllTools (maybe with some prefilteting step) and it works nicely.

Actually, I found another way to solve this problem by forked this adapter project.

The only modification is that I added a new dependency @n8n/json-schema-to-zod, and converted the json-schema to zod schema like below before I passed the params to DynamicStructuredTool

It worked well for me. And I also found that event we do not use DynamicStructuredTool to build the tool instance, we can directly use the tool() function , because we have got the zod-schema by this package.

And also, I found @n8n/json-schema-to-zod is the best package to solve this problem (I have tried several other packages).

We can find that the params of the function calling are correct now.

[postbird]

So I fixed my issues by down grading version a bit. Also am binding the whole getAllTools (maybe with some prefilteting step) and it works nicely.

Actually, I found another way to solve this problem by forked this adapter project.

The only modification is that I added a new dependency @n8n/json-schema-to-zod, and converted the json-schema to zod schema like below before I passed the params to DynamicStructuredTool

It worked well for me. And I also found that event we do not use DynamicStructuredTool to build the tool instance, we can directly use the tool() function , because we have got the zod-schema by this package.

And also, I found @n8n/json-schema-to-zod is the best package to solve this problem (I have tried several other packages).

We can find that the params of the function calling are correct now.

Hi, @benjamincburns, thank you for providing the this adopter for developers to integrate the mcp servers easier. I'm wondering if there are any issues of DynamicStructuredTool with the version of langchain?

[benjamincburns]

This has been fixed as of v0.4.1. Please update to the latest versions of @langchain/core, @langchain/mcp-adapters and whichever chat model provider library you're using e.g. @langchain/openai.

If you're still experiencing issues after updating to latest versions of ALL of the above libraries, please open a new issue.