Loading only some tools from an MCP server

[PaulRBerg]

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

MCP server tool descriptions gobble up the token context window in AI agent frameworks like Claude Code.

For instance, the GitHub MCP server alone increases the starting context window by ~46k tokens. So if I add this MCP server to my user scope (in ~/.claude.json), 25% of my context window will be gone as soon as I start a new chat (since Claude's max context window is 200k).

GitHub MCP has 91 tools, many of which I don't need (e.g. mcp__github__get_code_scanning_alert).

Is there any way to selectively load only some tools from an MCP server?

If the answer is no, would you be open to modifying the MCP spec to allow for partial loading of tools?

[natesilva]

I’d like to add another use case for this: Some MCPs are linked to paid services, and the MCP includes a combination of tools that are available to free and paid users, or even to different tiers of paid users. Without being able to selectively choose the tools that are loaded, an LLM may try to use one of the tools that is not available to my account level.

An example is the Brave search engine official MCP: https://github.com/brave/brave-search-mcp-server. It includes a tool called brave_summarizer that is only available with a Pro subscription, but this tool is enabled for any client.

Finally, as an example of a workaround, LM Studio’s MCP support allows turning off specific tools. Claude Desktop has a similar feature. I believe this is what OP is asking for; but it should be a core capability of MCP instead of an application-specific workaround.

LM Studio MCP configuration

Claude Desktop MCP configuration

[PaulRBerg]

I believe this is what OP is asking for; but it should be a core capability of MCP instead of an application-specific workaround.

indeed.

Without a core capability, we cannot selectively load tools in low-level environmenta like Claude Code.

[pugafran]

In my case, as a C# developer, I use the ModelContextProtocol methods to load all tools into a Semantic Kernel plugin. After that, I iterate through each tool and select only those that were chosen in the frontend. Then, I create a copy of the plugin containing only the selected tools.

(Semantic Kernel is a framework designed to integrate and orchestrate LLMs in applications. In Semantic Kernel terms, the MCPs are equivalent to plugins, and the functions are the tools)

[PaulRBerg]

Related issue in Claude Code repo: Feature: Declarative Tool Filtering in MCP Server Configuration