Glimind // privacy

Glimind is designed so that contributing data cannot leak your inputs.

From SDK telemetry we store ONLY:

• the tool / endpoint identifier
• a salted hash of the input shape (structure only) and an optional value-free skeleton
• a value-free output shape skeleton (structure of what a tool returns — never the values)
• success / failure and a normalized failure class (timeout, auth, rate-limit, schema, …)
• a normalized, scrubbed error signature (emails, IPs, tokens, long strings removed)
• latency and timestamp
• an optional coarse region (e.g. us, eu)
• coarse caller context: framework, model family, and language (e.g. langchain, claude, ts) — never identity
• a salted session id + call sequence, so we can learn tool-call order (the workflow graph) — the id is one-way hashed
• for the task you describe: a salted hash and a semantic embedding of the scrubbed phrase — used to learn which tools fit which tasks. The raw phrase is never stored; only the hash + vector survive.

We never store

• raw arguments, payloads, or tool outputs
• raw prompts or the raw text of your task/intent
• your reporter or session id in raw form (both are salted-hashed)
• any personal or identifying information

The ingest endpoint hard-rejects any unexpected field or anything that resembles a raw payload; shape skeletons are validated to contain only type tags. Everything we keep is structure or metadata — designed so the dataset is valuable to the whole ecosystem yet can never leak your data. See the SDK README for the exact client-side derivation.

Web analytics

For requests to our own site/API we keep aggregate counts only (by traffic source, page, region, and referrer) plus a salted, one-way hash of IP + user-agent to count unique vs returning visitors. We do not store raw IP addresses, and the hash cannot be reversed to identify you. These visitor hashes are retained for at most 90 days, then deleted.

Lawful basis & your rights

We process this structure/metadata under legitimate interests (operating and improving a neutral reliability service) with data minimisation by design — no raw payloads or PII are ever stored. You may request access to, or deletion of, any data associated with you (e.g. a reporter id or API key): email [email protected] and we act promptly. You can stop contributing at any time by uninstalling the SDK; server operators can remove a server from measurement at https://glimind.com/opt-out.

Retention

Value-free structure/metadata (shapes, error classes, aggregates, ecosystem history) is retained to keep the dataset useful over time. Anything closer to a person — visitor hashes — is capped at 90 days. We never sell personal data.

Probing

For servers we discover from public registries, we perform only the MCP initialize handshake and tools/list. We never invoke a tool. Probes carry a descriptive User-Agent linking here, honor a do-not-probe list, and are rate-limited. Opt out at https://glimind.com/opt-out.

Contact

Privacy questions or data requests: [email protected].