Your MCP servers spend context before you type a word
Every connected server sends its tool definitions to the model on each request. Separately, a coding agent with shell access can run commands you would never type yourself. These two tools measure the first problem and guard against the second.
The 200,000-token window, to scaleexample: 3,420 tokens =
Same sliver, magnified0 to 3,420
Two costs that stay invisible until you look
Context cost. MCP config files list servers, not tools, so nothing in your editor tells you what a server's tool definitions weigh. Tool names, descriptions and input schemas all count toward the model's context, and a server with a dozen verbose tools can outweigh the rest of your setup combined.
Guardrail risk. Claude Code can run shell commands and write files. A single git reset --hard, a rm -rf aimed at the wrong path, or a pasted API key written into a source file is enough to ruin an afternoon. Hooks can stop these before they run, but someone has to write and test them.
MCP Context-Cost Profiler
$19 one-time
Reads your Claude Code, Cursor, Claude Desktop or generic MCP configs and reports tokens per server and per tool. Simulates deferred loading, disabling low-use servers and shortening descriptions. An optional --verify flag checks its estimates against Anthropic's token counter.
Hook Guard
$19 one-time
Five tested Claude Code hook entries: block destructive shell commands, block secret-shaped strings and sensitive files, format edited files with formatters you already have, and keep a redacted JSONL log. Installer supports --dry-run and --uninstall.
Both are free to use from GitHub in full. The paid tier is a supporter tier. Both together are $29.
Get a first number without installing anything
Enter each server's name, how many tools it has and roughly how long its descriptions are. The calculator estimates the total, ranks your servers by cost and shows what deferring, disabling and shortening could save. It also states how wrong it can be. For a real per-tool audit of your actual configs, use the Profiler.