Can MCP be used without Claude?

[saurabhdea]

Pre-submission Checklist

• I have checked that this question would not be more appropriate as an issue in a specific repository
• I have searched existing discussions and documentation for answers

Question Category

• Protocol Specification
• SDK Usage
• Server Implementation
• General Implementation
• Documentation
• Other

Your Question

Is there a way to run MCP without the Claude API?

[tadasant]

Yes, MCP is not tied to Claude or the Anthropic API. Here are a few examples of using MCP without Claude:

• Use Gemini voice to manage Notion via Systemprompt
• Turn Python codebase to knowledge graph using cognee in Cline for VSCode
• Search the web while using Cursor
• Run Perplexity web searches with Sage

[Zhaobudaoyuema]

@tadasant Thank you for showing me. The information you've collected is really extensive.

[the-vampiire]

thank you for answering @tadasant. like OP, i thought only claude was capable. your examples look like MCP clients that use claude as their LLM backend. sorry if the terminology is wrong, i am still learning.

my mental model sees the following characters:

• LLM: the AI backend (claude, others)
• MCP client: a CL/GUI that supports MCP and calls out to the LLM and MCP server on user requests (or user permission for agentic workflow)
• MCP server: a local or hosted server that the MCP client interacts with to add external functionality to the LLM

this must be a dumb or x/y question because i cant find the answer anywhere: how do you integrate MCP understanding and usage into an LLM?

like how does an LLM know about the MCP servers available? is there something about claude training that makes it more "native" to it and can that be fine tuned into other LLMs?

again, sorry for the ignorance. if theres some reading you can RTFM kick me towards i wont be offended.

[tadasant]

@the-vampiire I don't think there is a centralized document that gets into this topic, so I'll just respond to each of your points specifically.

your examples look like MCP clients that use claude as their LLM backend.

The Use Gemini voice to manage Notion via Systemprompt example is not: that is powered entirely by Google Gemini, doesn't touch Anthropic's services at all.

Many MCP clients (e.g. Cursor, Cline, etc etc) offer a choice for what LLM you want to use as you use the app. So even the examples I gave for Cursor and Cline can avoid hitting Anthropic's services entirely (just so happens in the demo videos I was also using Claude).

You're right that most examples of MCP today involve Claude Sonnet 3.5 as the LLM powering them, but as evidenced by this^ it's entirely optional.

this must be a dumb or x/y question because i cant find the answer anywhere: how do you integrate MCP understanding and usage into an LLM?

The simple answer is that tool definitions are stuffed into your prompt as you click "enter". Like this.

is there something about claude training that makes it more "native" to it and can that be fine tuned into other LLMs?

There is something to this - in general Claude Sonnet is perceived by the community as being "better" at tool calling in this way than most other models. But other models can do it too, and I'm sure fine tuning could improve any model's capability in this regard.

[the-vampiire]

thanks a lot for the clarification. so is it a subject of some models (like claude) being better at tool calling in general or is there something specific about calling MCP tools that needs to be taught?

[tadasant]

If I understand your question correctly -- no, there should be no meaningful difference in a model's ability to do tool calls vs. MCP tool calls. MCP tool calls are just a particularly standardized form of tool call.