Is there a way to disable the human input tool

[rahuldave]

I find that even if I set the human_input_callback in Agent to None, the Agent gets set up with at-least one tool, the human input tool, and the llm (claude) prompts to use it. In my use case, I might do agentic RAG on 2 documents, but then want to just do a pure llm call with history to compare these two, or summarize with contrast. I use Augmented Claude for the first two, but then , on the third AugmentedClaude call, without a human input callback, the llm still requests for input because it has the tool. Is there a way of not exposing the human input tool to the llm?

I love the project! Have written a nice UI to talk with legal docs using it, and use the Meilisearch MCP tool to do a lot with full-text search ..and its so good i havent even needed embeddings as yet! I do intent to bring back human input, it's just that i dont want it at all times, and I've written a (for now) buggy websocket based callback. for it...

[saqadri]

I find that even if I set the human_input_callback in Agent to None, the Agent gets set up with at-least one tool, the human input tool, and the llm (claude) prompts to use it. In my use case, I might do agentic RAG on 2 documents, but then want to just do a pure llm call with history to compare these two, or summarize with contrast. I use Augmented Claude for the first two, but then , on the third AugmentedClaude call, without a human input callback, the llm still requests for input because it has the tool. Is there a way of not exposing the human input tool to the llm?

I love the project! Have written a nice UI to talk with legal docs using it, and use the Meilisearch MCP tool to do a lot with full-text search ..and its so good i havent even needed embeddings as yet! I do intent to bring back human input, it's just that i dont want it at all times, and I've written a (for now) buggy websocket based callback. for it...

@rahuldave thanks for the feedback and I was just running into this myself! For now, please set human_input_feedback in the MCPApp constructor to None. https://github.com/lastmile-ai/mcp-agent/blob/main/src/mcp_agent/app.py#L46

I will submit a fix to make that None by default. I am working on revamping human input anyway.

[rahuldave]

Super. I'd tried aetting to an empty array! Will use None. I also found telling Claude, "Do not ask for input' in the system prompt helped.