mcpfoldCurate tools, cut the context-window tax
Every MCP server dumps its full tool schema into the model's context whether you use those tools or not. Allow- or deny-list tools per server and only what you need loads — the committed benchmark cuts 45 tools to 9 for about 80% fewer tokens.
A model can only reason over what fits in its context window, and that budget is shared with the schema of every tool the connected servers expose. In the published fixture — GitHub, Supabase, and Playwright — that is 45 tools costing roughly 7,476 tokens before you have done any work.
mcpfold's proxy lets you allow- or deny-list tools per server, so only the handful you actually use is advertised. Keeping three tools per server in that fixture drops it to 9 tools and about 1,497 tokens — roughly 80% smaller. The number on this page is computed from the same committed benchmark methodology the docs publish, so the site and the benchmark can never disagree.
Keep only the tools you use
{
"servers": {
"github": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-github"],
"tools": { "allow": ["search_issues", "get_file_contents", "create_pull_request"] }
}
}
}Go deeper
Related features
Ready to try it? Install mcpfold — the CLI is free and MIT-licensed. Or head back to the overview.