# mcpcost > Health and token cost of every remote server in the official MCP registry, and of its npm and PyPI package servers. We probe each server each day. The site answers two questions for 40 000+ servers: Is this MCP server alive? How many context tokens does it cost? Token counts: claudeTokens is exact. It comes from the Anthropic count_tokens endpoint for claude-opus-5-5: one request with the server's tools plus a marker tool, one with the marker only, the difference is the cost of the tools. claudeTokensClaudeCode is the same with the Claude Code prefix mcp____ on each tool name. toolsTokens is an estimate: compact JSON of each tool (name, description, input_schema) counted with o200k_base. The exact count is about 1.9 times the estimate at the median. A server that needs auth may carry derived_* fields: the tool set of the publisher's open-source package, measured in a container; measured_with_credentials marks a server probed with stored credentials. A tool-use system prompt (286 tokens on Claude Opus 5.5) is added one time per request and is not in any number. ## Pages - https://mcpcost.ulehla.dev/ : overview and a live check form - https://mcpcost.ulehla.dev/tokens : how many context tokens an MCP server costs, with current percentiles - Package (stdio) servers from npm and PyPI are probed too, in a container with the network off; their page shows the npx or uvx command. - https://mcpcost.ulehla.dev/s/.xml : Atom feed of one server, an entry for every probe that changed its result, tool count or token cost - https://mcpcost.ulehla.dev/changes : what changed since the previous run (new, removed, down, back, tool sets that grew or shrank); Atom feed at /changes.xml - https://mcpcost.ulehla.dev/s/ : one server, for example https://mcpcost.ulehla.dev/s/com.replynodes/mcp - https://mcpcost.ulehla.dev/servers?q=&status=ok&sort=claude : search and filters (sort=tokens for the estimate) - https://mcpcost.ulehla.dev/check?target= : probe a server that is not in the registry; a form there tracks it on the board (name url//) - https://mcpcost.ulehla.dev/watch?server= : e-mail alerts when a server stops answering, comes back, or its tool set or token cost changes - https://mcpcost.ulehla.dev/report : monthly reports, The state of the MCP registry; https://mcpcost.ulehla.dev/report/2026-10 reads the runs of October 2026 and ends with a CC BY 4.0 dataset, one row per server, at https://mcpcost.ulehla.dev/data/mcpcost-2026-10.json.gz - https://mcpcost.ulehla.dev/statistics, /hosts, /methodology ## Machine access - JSON, one server: https://mcpcost.ulehla.dev/api/s/.json - JSON, search: https://mcpcost.ulehla.dev/api/servers.json?q=&status=ok&sort=claude&page=1 (at most 6 requests in 10 seconds from one address; answers are cached for three hours, so repeat a query freely) - JSON, the board's row for an endpoint URL, no probe: https://mcpcost.ulehla.dev/api/lookup.json?url= (302 to the server's JSON, 404 when the board does not know the URL) - Compare two to five servers side by side: https://mcpcost.ulehla.dev/compare?s=&s= - JSON, live probe: https://mcpcost.ulehla.dev/api/check.json?url= - JSON, run summary: https://mcpcost.ulehla.dev/api/summary.json - Re-check evidence for the registry, the newest run, in the row shape of modelcontextprotocol/registry#1715 (entry, checked_at, check, state, evidence; gzip of JSON, CC BY 4.0, changes after each daily run): https://mcpcost.ulehla.dev/data/mcpcost-evidence-latest.json.gz ; its index (run date, rows, size, SHA-256): https://mcpcost.ulehla.dev/data/mcpcost-evidence-latest.index.json - JSON, servers tracked outside the registry: https://mcpcost.ulehla.dev/api/tracked.json - MCP server (Streamable HTTP, no auth): https://mcpcost.ulehla.dev/mcp with tools lookup_server, search_servers, compare_servers (two to five side by side, as JSON, Markdown or CSV), check_server, registry_summary - CLI for your own config: npx mcpcost (https://www.npmjs.com/package/mcpcost) - Source and raw data: https://github.com/ondraulehla/mcpcost (MIT)