tool from an MCP server gets called too many times

[dmigra0]

What happened?

When an LLM instructs to call an MCP tool, the tool gets called too many times.
This is mostly reproducibly by using gpt-oss model in ollama

Version Information

33312fb40974e78dc4f32d77e7f204ef14f6b9f

Steps to Reproduce

1. Get an MCP server

git clone https://github.com/jonigl/mcp-server-with-streamable-http-example.git
cd mcp-server-with-streamable-http-example
uv run mcp-server

3. Configure an MCP server e.g:```
   streamable:
   type: streamable-http
   url: http://127.0.0.1:8000/mcp

3. Select the server in a chat 
4. Ask a question: "generate a random number using streamble then pass it to hello_world"
5. What you shall observe is that random_number gets called multiple times, sometimes it hits recursion level limit, sometimes it fails to pass results to next call. Sometimes it forgets what the question was

### What browsers are you seeing the problem on?

Firefox

### Relevant log output

```shell
2025-09-22T07:30:41.569Z debug: [BaseClient] userMessage
{
  messageId: "b43826ef-cfa3-4fa7-90a0-03e4cac060b5",
  parentMessageId: "c3f93cd8-eb75-4c96-b996-609a59a162fe",
  conversationId: "260a6e6f-7086-44b0-84cb-2d96336d8511",
  sender: "User",
  text: "generate a random number using streamble then pass it to hello_world",
  isCreatedByUser: true,
  tokenCount: 17,
}
2025-09-22T07:30:41.577Z debug: [saveConvo] api/app/clients/BaseClient.js - saveMessageToDatabase #saveConvo
2025-09-22T07:30:49.664Z debug: [MCP][streamable] Transport sending: {"method":"ping","jsonrpc":"2.0","id":10}
2025-09-22T07:30:49.668Z debug: [MCP][streamable] Transport received: {"jsonrpc":"2.0","id":10,"result":{}}
2025-09-22T07:30:49.669Z debug: [MCP][streamable] Transport sending: {"method":"tools/call","params":{"name":"random_number","arguments":{"min_val":1,"max_val":100}},"jsonrpc":"2.0",... [truncated]
2025-09-22T07:30:49.673Z debug: [MCP][streamable] Transport received: {"jsonrpc":"2.0","id":11,"result":{"content":[{"type":"text","text":"38"}],"structuredContent":{"result":38},"is... [truncated]
2025-09-22T07:30:49.673Z debug: [MCP][User: 68bec54c66e9ee02dcb99cf0] Updated last activity timestamp: 2025-09-22T07:30:49.673Z
2025-09-22T07:30:51.582Z debug: [MCP][streamable] Transport sending: {"method":"tools/call","params":{"name":"random_number","arguments":{"min_val":1,"max_val":100}},"jsonrpc":"2.0",... [truncated]
2025-09-22T07:30:51.588Z debug: [MCP][streamable] Transport received: {"jsonrpc":"2.0","id":12,"result":{"content":[{"type":"text","text":"63"}],"structuredContent":{"result":63},"is... [truncated]
2025-09-22T07:30:51.589Z debug: [MCP][User: 68bec54c66e9ee02dcb99cf0] Updated last activity timestamp: 2025-09-22T07:30:51.589Z
2025-09-22T07:30:55.234Z debug: [FinalizationRegistry] Cleaning up client for user 68bec54c66e9ee02dcb99cf0
2025-09-22T07:30:55.234Z debug: [FinalizationRegistry] Cleaning up client for user 68bec54c66e9ee02dcb99cf0
2025-09-22T07:30:55.234Z debug: [FinalizationRegistry] Cleaning up client for user 68bec54c66e9ee02dcb99cf0
2025-09-22T07:31:03.349Z debug: [MCP][streamable] Transport sending: {"method":"tools/call","params":{"name":"random_number","arguments":{"min_val":1,"max_val":100}},"jsonrpc":"2.0",... [truncated]
2025-09-22T07:31:03.355Z debug: [MCP][streamable] Transport received: {"jsonrpc":"2.0","id":13,"result":{"content":[{"type":"text","text":"65"}],"structuredContent":{"result":65},"is... [truncated]
2025-09-22T07:31:03.355Z debug: [MCP][User: 68bec54c66e9ee02dcb99cf0] Updated last activity timestamp: 2025-09-22T07:31:03.355Z
2025-09-22T07:31:06.117Z debug: [MCP][streamable] Transport sending: {"method":"tools/call","params":{"name":"random_number","arguments":{"min_val":1,"max_val":100}},"jsonrpc":"2.0",... [truncated]
2025-09-22T07:31:06.122Z debug: [MCP][streamable] Transport received: {"jsonrpc":"2.0","id":14,"result":{"content":[{"type":"text","text":"35"}],"structuredContent":{"result":35},"is... [truncated]
2025-09-22T07:31:06.122Z debug: [MCP][User: 68bec54c66e9ee02dcb99cf0] Updated last activity timestamp: 2025-09-22T07:31:06.122Z
2025-09-22T07:31:11.156Z debug: [MCP][streamable] Transport sending: {"method":"tools/call","params":{"name":"random_number","arguments":{"min_val":1,"max_val":100}},"jsonrpc":"2.0",... [truncated]
2025-09-22T07:31:11.160Z debug: [MCP][streamable] Transport received: {"jsonrpc":"2.0","id":15,"result":{"content":[{"type":"text","text":"44"}],"structuredContent":{"result":44},"is... [truncated]
2025-09-22T07:31:11.160Z debug: [MCP][User: 68bec54c66e9ee02dcb99cf0] Updated last activity timestamp: 2025-09-22T07:31:11.160Z
2025-09-22T07:31:13.567Z debug: [MCP][streamable] Transport sending: {"method":"tools/call","params":{"name":"hello_world","arguments":{"name":"44"}},"jsonrpc":"2.0","id":16}
2025-09-22T07:31:13.572Z debug: [MCP][streamable] Transport received: {"jsonrpc":"2.0","id":16,"result":{"content":[{"type":"text","text":"Hello, 44!"}],"structuredContent":{"result"... [truncated]
2025-09-22T07:31:13.572Z debug: [MCP][User: 68bec54c66e9ee02dcb99cf0] Updated last activity timestamp: 2025-09-22T07:31:13.572Z
2025-09-22T07:31:16.230Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 676,
  completionTokens: 749,
}
2025-09-22T07:31:16.230Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 685,
  completionTokens: 142,
}
2025-09-22T07:31:16.231Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 694,
  completionTokens: 1090,
}
2025-09-22T07:31:16.231Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 703,
  completionTokens: 211,
}
2025-09-22T07:31:16.231Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 712,
  completionTokens: 426,
}
2025-09-22T07:31:16.232Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 721,
  completionTokens: 152,
}
2025-09-22T07:31:16.232Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 734,
  completionTokens: 181,
}
2025-09-22T07:31:16.233Z debug: [spendTokens] conversationId: 260a6e6f-7086-44b0-84cb-2d96336d8511 | Context: message | Token usage:
{
  promptTokens: 961,
  completionTokens: 117,
}
2025-09-22T07:31:16.242Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.243Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.243Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.243Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.245Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.245Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.248Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.248Z debug: [spendTokens] No transactions incurred against balance
2025-09-22T07:31:16.252Z debug: [saveConvo] api/app/clients/BaseClient.js - saveMessageToDatabase #saveConvo
2025-09-22T07:31:16.257Z debug: [AgentController] Request closed
2025-09-22T07:31:16.259Z debug: [AgentController] Performing cleanup
2025-09-22T07:31:16.259Z debug: [AgentController] Cleaning up abort controller
2025-09-22T07:31:16.259Z debug: [AgentController] Cleanup completed
2025-09-22T07:31:35.972Z debug: [FinalizationRegistry] Cleaning up client for user 68bec54c66e9ee02dcb99cf0
2025-09-22T07:32:54.836Z info: Cleaning up FlowStateManager intervals...

Screenshots

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

This is an LLM issue and is common for smaller/local LLMs. In this case, extra prompting is warranted, recommended via agent instructions.

[ChrisAlthaus]

Do you have any recommendations for the extra prompting (system-prompt) messages? I'm having the same issue with gpt-oss:120b. I have tried to alternate the modell settings, like deactivate streaming and changing the temparature, but this seemingly does not lead to a good solution.
Also when I write in the system prompt that the same method should only be called once, the reasoning process knows this fact but nevertheless calls the method repeatedly. Then finally the user query or question is somehow forgotten and the model or librechat does not know what to do.

[JulianHandrup]

GPT-OSS seems to go overboard with function calls a lot. It has a lot of trouble calling the right tool and often tries to call the tool by the action name, not pass the action as a parameter to the tool and call the tool by its name (I set the tool up like the OpenWeather tool that's in the repo).
I'm having the same issue and found it useful to split the operations using two or three agents and the new agent handoff system where I use one agent running qwen3-code or similiar and one agent running gpt-oss basically doing the post-processing of the data.
You can add a third agent as a supervisor that delegates tasks between the agents available for handoffs.

Though, I haven't tried this in a use-case where I need multiple tool requests like "listPersons()" -> choose person -> "getPerson(id)" -> extract department id -> "getDepartment(id) for a prompt like "In which department does john smith work?".