Privacy

Privacy policy

Minerval collects the minimum it needs to run a metered claim-graph service, and nothing it collects is sold, shared for advertising, or used to track you anywhere else. What it does keep, including a contributor’s track record on the graph, is described below.

Effective September 2, 2026 · applies to minerval.ai, the Minerval API, and the Minerval browser extension.

Accounts

Signing in uses OAuth through GitHub or Google. We store the provider and account identifier (e.g. github:12345), your display name, and nothing else from your provider profile. We never see or store a password. API keys you create are stored only as salted hashes; the full key is shown once, to you.

Contributor standing

The graph is an open commons, and contributions to it are attributed: a contributor’s track record is part of the graph’s public record, the same way a cited source’s track record is. From those outcomes each account accrues standing, which exists to keep sincere participation free while spam and deliberate manipulation carry a cost. Standing governs what an account may contribute, and because the graph weighs every source by its reliability, a contributor’s record may likewise inform how much weight their submissions carry. It is built solely from contributions you choose to make; reading claims, analyzing pages with the extension, and API usage never affect it. Changes to standing are recorded with the contribution that caused them, and adverse outcomes can be appealed. Standing is never sold or used for advertising.

Usage metering

Requests that invoke Minerval’s analysis agents are metered against your account: for each call we record which agent ran, the model, token counts, and derived cost. This is billing and rate-limiting bookkeeping; it does not include the content you analyzed. Reading claims is unauthenticated and unmetered, and read requests are not tied to any account.

Agent transcripts

The graph’s own work is kept on the record. When you submit a URL for ingestion, the agents that read it and reason about it leave a transcript: the fetched document, their searches, and their reasoning, attributed to the submitting account and kept for thirty days so the process can be inspected and evaluated. Sources are public pages by construction (our servers fetch them without credentials), and the claims they yield become part of the public graph with their provenance.

Content you hand over to be analyzed rather than ingested is different, and is never part of a transcript: the page open in your browser when you use the extension, questions typed into its chat, and passages sent to the MCP server’s analysis tools are processed and then discarded. Only the metering record described above remains.

The browser extension

The extension sends a page’s readable text, its URL, and its title to the Minerval API only when you ask it to: by pressing Analyze page in the popup, or by opting a site (or all sites) into automatic analysis in the extension’s settings. Nothing is sent by default, and any site can be disabled entirely. Questions you type into the popup chat, along with the page’s claim context, are likewise sent only when you send them.

Sent page text is used for exactly one purpose: extracting the claims on the page and matching them against the graph so the extension can annotate what you are reading and answer your questions. Page text is transient: it is never written to our database or to agent transcripts, and no claim is added to the graph from it. Analysis results are cached in server memory, keyed by the page URL and a hash of its content, so re-analyzing an unchanged page is instant and is not re-processed; the cache expires within about an hour and does not survive a server restart. Chat questions are handled the same way and are not retained at all. Analysis requests authenticate with your API key and are metered as described above.

The extension stores your settings (API key, markup level, and per-site preferences) in your browser’s extension storage (synced by your browser if you use profile sync). It sets no cookies, injects no trackers, runs no analytics, and never rewrites the text of pages you visit; annotations are drawn as an overlay.

Service providers

Analysis runs on infrastructure and model providers acting on our behalf: pages you submit are processed by third-party large language model providers (currently Anthropic, OpenAI for embeddings, and DeepSeek routed through OpenRouter), always under API terms that exclude training on your data, and our servers run on AWS with the web app hosted on Vercel. These providers process data to provide the service and for no other purpose.

Retention & deletion

Cached page analyses expire from server memory within about an hour; agent transcripts from source ingestion are deleted after thirty days; metering records are retained as account history. To delete your account and its associated records, email us and we will remove them.

Contact

Questions or deletion requests: privacy@minerval.ai. See also about for who runs Minerval.