connect

Ask your assistant

mcpcost is an MCP server too. Add it to your assistant and ask if an MCP server is up, how a client gets access and how many context tokens its tools cost. Five tools, all read-only, no sign-in, free.

Add it

The server speaks Streamable HTTP and needs no authentication. The URL:

https://mcpcost.ulehla.dev/mcp

Claude: mcpcost is in the Claude directory: open its listing and connect it there, or find mcpcost under Customize, Connectors. No sign-in is asked. To add it by hand, choose “Add custom connector” and give the URL above.
Claude Code: claude mcp add --transport http mcpcost https://mcpcost.ulehla.dev/mcp
ChatGPT: In developer mode, add a custom connector (app) with the URL above and no authentication.
Other clients: Cursor, VS Code, Windsurf and other clients that take a remote server: add an entry with the URL above to the MCP config of the client.
Local: npx -y mcpcost mcp starts a stdio server with the four tools that read the board, and with two tools that measure a server or your own config on your machine. The package on npm.

The tools, and what to ask

lookup_server: Look up a registry server. The newest daily probe of one server: status, how a client gets access, the exact token cost of its tools in Claude, the tool count, the latency and its most expensive tools.
Ask: “Is the GitHub MCP server up, and how many tokens do its tools cost?”
search_servers: Search the registry board. Servers by words in their name, title, description or host, with filters, sorted by token cost.
Ask: “Find MCP servers for Postgres that answer without auth, the cheapest first.”
compare_servers: Compare servers side by side. Two to five servers next to each other: status, access, tools, token cost, share of a 200k context window.
Ask: “Compare the token cost of the Notion, Linear and Atlassian MCP servers.”
check_server: Probe a server live. Connects to a public MCP URL now, with no credentials, and reports status, protocol, tools and token cost.
Ask: “Check https://example.com/mcp: does it answer, and what do its tools cost?”
registry_summary: Registry summary. The totals of the newest run: how many servers answer and why the others do not, token cost percentiles, auth, protocols, the biggest publishers.
Ask: “How many servers in the MCP registry answer today, and what does the median server cost?”

Where the answers come from

The board probes each server of the official MCP registry once a day: remote servers over the network, npm and PyPI package servers in a container. Token counts for Claude are exact, from the Anthropic count_tokens endpoint. Four tools read that data and connect to nothing. check_server connects to the URL you give, with no credentials, and waits 10 seconds at most. How the probe works.

Limits: Only public https URLs. No tool changes data, sends a message or spends money. Answers are JSON of a few kilobytes; a search page holds 20 servers at most.
Privacy: The server gets the arguments of a tool call and nothing else. It keeps nothing in its database. Privacy policy.
Support: Open an issue at github.com/ondraulehla/mcpcost. The source is there too, under the MIT licence.