Tool metadata is always in context
Every tool's name, description, and JSON schema is sent to the model on every request — before the user even types. A server with 50 verbose tools can quietly cost tens of thousands of tokens each call.
Free MCP inspector
See what an MCP server really costs your context window.
Paste an MCP server URL to inspect its tools, resources, and prompts — and the exact token footprint they add to every request.
Try:
Why it matters
MCP tools eat context before you type a word.
A few tools dominate the budget
Token cost is rarely even. One or two tools with deep, nested input schemas often account for most of a server's footprint. We rank them so you know exactly what to trim.
Inspect every tool, resource, and prompt
Search the full tool list, expand any tool to see its arguments and schema size, and read each prompt's actual rendered content — all before you connect the server to an agent.
Frequently asked questions
What is an MCP server's context cost?
MCP (Model Context Protocol) servers expose tools, resources, and prompts to an AI agent. The metadata describing those tools — names, descriptions, and input schemas — is injected into the model's context window on every request. That metadata consumes tokens whether or not a tool is ever called, so a large server can meaningfully shrink the room left for your actual prompt.
How are the token counts calculated?
We serialize each tool, resource, and prompt definition the way a model receives it and count tokens with the o200k_base tokenizer. It's an estimate: exact counts vary slightly between model families, but it's an accurate way to compare servers and spot expensive tools.
Do you store the URL or my token?
No. The scan runs server-side for a single request. Tokens and custom headers you enter under Advanced are used only to fetch that one server and are never stored or logged.
Which servers can I scan?
Any MCP server reachable over HTTP that speaks the Streamable HTTP transport. Public servers work out of the box; for protected ones (like GitHub's), add a bearer token under Advanced options. For servers running on your machine, you can use something like ngrok to expose it to the internet, and then paste the ngok url.
Building agents that actually use these tools?
TeamCopilot lets you build AI automations in plain English — connect MCP servers and other tools, keep your team in the approval loop, and run everything on your own cloud.