Remote data handling#

This policy applies when using the hosted remote MCP. The local PyPI/stdio package has its own data handling.

Requests are processed to execute read-only MCP tools. Relevant search text, names, aliases and QIDs are sent to Wikimedia's Wikidata services. Caller-supplied websites are compared as identity evidence; they are not fetched. Public remote V1 does not contact Google Knowledge Graph or any model provider and requires no provider secrets or sign-in. Local installations retain optional Google support.

Request content is not intentionally retained, published, turned into a public corpus or stored in a resolver cache. The service does not intentionally log full tool arguments, prompts, names, addresses, URLs or coordinates. Native Worker logs, traces and Logpush are disabled. Do not send credentials in tool arguments.

Operational storage contains the global daily/minute admission counters and aggregates by day and tool: calls/errors, provider attempts/errors, response bytes and latency histogram buckets. Completion metrics are best effort. Aggregates are pruned to the current UTC day and preceding 30 days by both requests and a persistent daily alarm. No per-request identity records are stored. The quota row holds the current day only. A temporary isolate cache remembers exhaustion until midnight UTC.

A coarse Cloudflare burst limiter uses the connecting network address as an ephemeral rate-limit key. Application code does not store it or treat it as a person's identity. Cloudflare processes network traffic and supplies aggregate runtime metrics under its privacy policy. This policy does not promise that Cloudflare keeps no infrastructure/security records. Wikidata's own infrastructure receives the relevant queries.

For questions use the project support page. Report sensitive security issues through the confidential reporting instructions.