Your MCP servers are eating your context window. I built a CLI to audit the tax

Every time you launch a new MCP server, it injects all of its tool schemas into every request. Claude Code loads these schemas at the beginning of a session, and there's no UI to temporarily disable a server you're not using at that moment. Some servers can consume up to 16.3% of a 200k context window before the conversation even begins.

To address this issue, I created a simple command-line interface called mcp-tax. This tool audits the context tax of your MCP servers, allowing you to launch Claude Code with the most expensive servers temporarily disabled - per session, without altering your actual configuration. It's written in Python using only the standard library, so there are no dependencies.

To use it, simply install with pip and run mcp-tax list to see which servers are configured and which are disabled. The tool will then show you the estimated number of tokens each server is consuming, based on the size of its schema. You can then toggle servers on or off using mcp-tax off and mcp-tax on, with the changes persisting in ~/.config/mcp-tax/disabled.json.

When you're ready to launch Claude Code, use mcp-tax run to execute it with a filtered configuration that excludes the disabled servers. This provides an order-of-magnitude gauge of which server might be eating up your context window, rather than a precise billing meter. Note that the tool only works with stdio servers and not with SSE or streamable HTTP servers.

It also doesn't account for runtime behavior, so even a server with a small schema could still be expensive if its tool results are large.