How do I trigger MCPs?

[vmajor]

In Claude Desktop they trigger autonomously as needed, in my own GUI I can trigger them with custom prompts that I inject depending on which MCPs I enable, but before I start tearing LibreChat apart, what is the correct way to let the LLM know that MCPs exist. I have these enabled right now and none can be triggered:

LibreChat         | 2025-03-11 06:15:03 info: [MCP] Initialized 4/4 server(s)
LibreChat         | 2025-03-11 06:15:03 info: [MCP][puppeteer] ✓ Initialized
LibreChat         | 2025-03-11 06:15:03 info: [MCP][filesystem] ✓ Initialized
LibreChat         | 2025-03-11 06:15:03 info: [MCP][brave-search] ✓ Initialized
LibreChat         | 2025-03-11 06:15:03 info: [MCP][search1api] ✓ Initialized
LibreChat         | 2025-03-11 06:15:03 info: [MCP] All servers initialized successfully

[aallsbury]

Having same problem here

[danny-avila]

https://www.librechat.ai/docs/features/agents#model-context-protocol-mcp

[vmajor]

OK thank you @danny-avila

This kind of works, but with llama-server I have this issue:

got exception: {"code":500,"message":"Cannot use tools with stream","type":"server_error"}
srv  log_server_r: request: POST /v1/chat/completions 127.0.0.1 500

Unfortunately llama-server itself does not seem to have a startup flag that disables streaming, it can be passed as an argument during a request, but there is no option to include this in the request sent to the API, not as far as I can see.

Can/how can stream=false be included in the custom endpoint definition in librechat.yaml?

[Drakosfire]

@vmajor Hah! I have been tearing the code apart, I have added logging and tracked down as many places as I could to try and work around this!
I can't find it!

From my exploration the Agent client extension of Base client is using processStream. So where in there stream=false is being overwritten in the model options.

@danny-avila Am I understanding correctly that by using Agents it is forced into streaming tokens?

Ie, is this a dead end and I can't use llama.cpp with as my local inference engine for agents and tool calling?