I am facing a similar issue with while trying to complete the unit 2 of HF mcp-course Building a Tiny Agent with TypeScript

Borrowing from your json config, I use

{
	"model": "llama3.2:3b",
	"endpointUrl": "http://localhost:11434",
	"servers": [
		{
			"type": "stdio",
			"config": {
				"command": "npx",
				"args": [
					"mcp-remote",
					"http://localhost:7860/gradio_api/mcp/sse"
				]
			}
		}
	]
}

and running npx @huggingface/tiny-agents run ./my-agent

$ npx @huggingface/tiny-agents run ./my-agent
[189887] Using automatically selected callback port: 19650
[189887] [189887] Connecting to remote server: http://localhost:7860/gradio_api/mcp/sse
[189887] Using transport strategy: http-first
[189887] Received error: Error POSTing to endpoint (HTTP 405): Method Not Allowed
[189887] Recursively reconnecting for reason: falling-back-to-alternate-transport
[189887] [189887] Connecting to remote server: http://localhost:7860/gradio_api/mcp/sse
[189887] Using transport strategy: sse-only
[189887] Connected to remote server using SSEClientTransport
[189887] Local STDIO server running
[189887] Proxy established successfully between local STDIO and remote SSEClientTransport
[189887] Press Ctrl+C to exit
[189887] [Local→Remote] initialize
[189887] {
  "jsonrpc": "2.0",
  "id": 0,
  "method": "initialize",
  "params": {
    "protocolVersion": "2025-03-26",
    "capabilities": {},
    "clientInfo": {
      "name": "@huggingface/mcp-client (via mcp-remote 0.1.9)",
      "version": "0.2.0"
    }
  }
}
[189887] [Remote→Local] 0
[189887] [Local→Remote] notifications/initialized
[189887] [Local→Remote] tools/list
[189887] [Remote→Local] 1
Agent loaded with 1 tools:
- predict
> analyse this sentence: I am feeling good today!
<Tool call_5v1bkkdp>
predict {"text":"I am feeling good today!"}
/home/sylvain/node_modules/@huggingface/tiny-agents/dist/cli.js:120
    throw err;
    ^

SyntaxError: Unexpected non-whitespace character after JSON at position 35
    at JSON.parse (<anonymous>)
    at Agent.processSingleTurnWithTools (/home/sylvain/node_modules/@huggingface/mcp-client/dist/src/index.js:233:71)
    at process.processTicksAndRejections (node:internal/process/task_queues:95:5)
    at async Agent.run (/home/sylvain/node_modules/@huggingface/mcp-client/dist/src/index.js:328:9)
    at async mainCliLoop (/home/sylvain/node_modules/@huggingface/tiny-agents/dist/cli.js:131:22)
    at async main (/home/sylvain/node_modules/@huggingface/tiny-agents/dist/cli.js:301:5)

Node.js v20.12.2
[189887] 
Shutting down...

leveraging the local model doesn't seem to be an issue here as we see a tool call with I guess a legit call to the mcp tool... but still a json.parse issue in the end...

@alexdconf my problem may be related to yours after all, inspecting .../node_modules/@huggingface/mcp-client/dist/src/index.js and putting console.log everywhere to debug, it seems that toolCall.function.arguments in line 233 contains twice the arguments string because of lines 222-224, thus commenting 222-224

// if (toolCall.function.arguments) {
//   finalToolCalls[toolCall.index].function.arguments += toolCall.function.arguments;
// }

fixes the parsing issue as we now have a well formed json dump.

I can make it work with the following example

npx @huggingface/tiny-agents run ./get-started

with the following agent config

{
  "model": "llama3.2:3b",
  "endpointUrl": "http://localhost:11434",
  "servers": [
    {
      "type": "http",
      "config": {
        "url": "https://evalstate-hf-mcp-server.hf.space/mcp"
      }
    }
  ]
}

and the prompt find details about MNIST dataset, however when the LLM pass an argument that should be a number as a string, I am getting an error, for instance

> search about llama3 model
<Tool call_igzqggqb>
model_search {"author":"null","library":"null","limit":"100","query":"llama3","sort":"downloads","task":"null"}[
  {
    id: 'call_igzqggqb',
    index: 0,
    type: 'function',
    function: {
      name: 'model_search',
      arguments: '{"author":"null","library":"null","limit":"100","query":"llama3","sort":"downloads","task":"null"}'
    }
  }
]

ERROR: {"author":"null","library":"null","limit":"100","query":"llama3","sort":"downloads","task":"null"}
JSON PARSE string

/home/sylvain/node_modules/@huggingface/tiny-agents/dist/cli.js:120
    throw err;
    ^

McpError: MCP error -32602: MCP error -32602: Invalid arguments for tool model_search: [
  {
    "code": "invalid_type",
    "expected": "number",
    "received": "string",
    "path": [
      "limit"
    ],
    "message": "Expected number, received string"
  }
]

Hi!

I think there are a couple of different issues here, but I think one way to debug is to feed back errors into the LLM. Let me try to open a PR to do this.

in the meantime, try using a larger model! It will generate valid JSON more reliably

If you are able, can you please give a try to #1511 locally?

Hello, thanks for following up on this, as far as my problem goes it seems that the size of the model is not the problem. I tested it with llama3.2:3b and qwen2.5-coder:7b, both model producing legit json.dump outputs ending with the same error.

The problem came from the doubling of the tool_call arguments which I overcame by commenting lines 222-224 of mcp-client/..../index.js

For the other issue concerning a tool expecting an argument as number, I guess that's a common issue when you deal with functions as mcp/llm tools, you'd better treat the inputs as str no matter what and deal with conversion to number or whatever in the tool core logic, otherwise you need to enforce a type checking before you call a tool...

I can give a try to #1511, although I am not quite sure how, as as soon as the error strike the tiny-agent exits.

we'll push a new release with #1511 inside, so it's easier to test it locally. (otherwise you have to build locally using pnpm)

ok ok @sdelahaies i can now reproduce your error. Let me see how to best fix it

Ok, how does this look @sdelahaies? #1512

Hello, it does solve the issue indeed, great, thanks!

thanks a lot for the debugging info + resolution proposal @sdelahaies, it's been merged