Security
What is stored, who can read it, and what the model is sent.
PatentReply gives your AI a compiled, source-linked record of a U.S. application. This page says what that involves, in the order a prosecutor asks: what is public, who has access, what leaves for the model, what is kept, and what has and has not been audited.
- 1A matter is compiled from the public USPTO record. Only what you add is private.
- 2Every connection is a signed-in person, through OAuth. There are no API keys.
- 3Every passage your AI reads carries its source address in the stored record.
- 4PatentReply does not train models on your matters. Your AI provider’s retention is theirs.
- 5Private data is deleted on request within 30 days. Compiled public records are kept.
- 6Built to the SOC 2 criteria, on providers that hold SOC 2 Type II reports.
What is public and what is private
A matter is compiled from the public USPTO record: the published application, its file wrapper, and the patents and published applications the examiner cites. Those are public documents, and most of what the model reads is one of them.
What you add is private to your account: documents you upload (a filed amendment, a non-patent reference), response strategies, proposed amendments, approvals, and exports of firm work product. Only published applications can be compiled.
Access
Every connection is a signed-in person. Your AI client authenticates through OAuth against PatentReply’s sign-in service (authorization code with PKCE, refresh tokens, bearer tokens only). There are no API keys to copy into a client, and nothing to leak from one.
Access from a client is revoked by removing the connector in that client. Sign out everywhere on your Settings page revokes every client at once. Approvals happen on patentreply.ai, signed in as you; a “yes” in a chat does not count, and the model cannot approve on your behalf.
Site served with HTTP Strict Transport Security · sign-in issuer published at /.well-known/oauth-authorization-server on the record service
What the model is sent
When your connected AI reads a matter through PatentReply, it asks for documents and receives passages with their source addresses; answers can be traced back to the stored bytes they quote, and a fact labelled verified_source was quote-matched against the document it came from.
Training and your AI provider
PatentReply does not train models on your matters, your uploads, your strategies, or your approvals, and does not use them to improve models for anyone else.
PatentReply supplies the record to the AI you connect. What that provider retains is governed by your agreement with them, not by PatentReply. A firm-managed or zero-retention plan with your provider is the right setting for client matters; a consumer account is not, and you should say so to the client.
PatentReply’s own indexing of the file wrapper calls model providers on API terms, which exclude training on the data sent. PDFs, including non-patent references you upload, are converted to text by Datalab. No prompt-tracing service is enabled in production.
Retention and deletion
Compiled public records are retained so a matter can be reopened without recompiling; removing a matter from your list keeps the compiled record. They are public documents.
Your private data (uploads, strategies, proposed amendments, approvals, exports) is deleted on request within 30 days. Write to [email protected] from your account email. Deleting your account deletes your private data on the same terms. Usage logs and analytics are kept for the periods in the privacy policy.
Encryption and where data lives
Data is encrypted in transit; the site enforces HTTPS with Strict Transport Security. At rest, the database and file storage are encrypted by the managed provider that holds them. Card details are handled by Stripe and never touch PatentReply.
These providers process data on PatentReply’s behalf:
- Supabase
- Database, file storage, and sign-in (the OAuth issuer your AI client signs in through)
- Vercel
- Hosts patentreply.ai; cookieless page-performance analytics; AI Gateway, which routes model requests
- Railway
- Hosts the record service your AI connects to (mcp.patentreply.ai) and its logs
- Cloudflare
- Edge network and TLS in front of the site
- Anthropic
- Identifies and indexes the cited references in a file wrapper at compile, on API terms
- TypeSafe
- Citation-reading model, reached through Vercel AI Gateway with zero data retention requested
- Datalab
- Converts PDFs to text: USPTO file-wrapper papers, and non-patent references you upload
- Stripe
- Payments and invoices; card details never reach PatentReply
- Resend
- Account and billing email
- PostHog
- Product analytics: usage events with account and application numbers, never matter content; no session recordings
- WorkOS
- Sign-in for firm service accounts only
SOC 2 and audits
PatentReply is built to the SOC 2 Trust Services Criteria for security, availability and confidentiality: every connection is an authenticated person, each account’s work is isolated from every other’s and that isolation is tested on every code change, data is encrypted in transit and at rest, and retention periods are enforced on a schedule.
The infrastructure and services PatentReply runs on hold SOC 2 Type II reports: Supabase, Vercel, Railway, Cloudflare, Anthropic, TypeSafe, Datalab, Stripe, Resend, PostHog and WorkOS.
PatentReply’s own SOC 2 report is not yet available. Security questionnaires are answered by email, and enterprise and zero-retention arrangements are available. Write to [email protected].
To report a vulnerability, write to the same address with “security” in the subject. We reply to every report.
What PatentReply does not do
It does not file with the USPTO, does not compute deadlines, and does not give legal advice. It is not a law firm; it helps AI work from the record and keeps legal judgment with the practitioner of record. Approvals and legal judgment
Statements checked 2026-10-02 · Questions to [email protected] · Privacy policy