Opyt
use cases docs install

Privacy policy

What Opyt stores, and what it never collects.

The short version: a local install phones nobody. If you use your library from claude.ai or ChatGPT, Opyt's server keeps a copy of it to serve back to you, and holds a change you ask for there until your own computer has made it. The sharing service counts what it serves in daily totals and publishes those totals. Nothing here is sold, mined across users, or used to train models.

Changed on 2026-11-18, and it stores more. A reader view is made on Opyt's server again, so it opens on any device and its comments are kept there, and your Mac keeps a copy of all your reader views and their comments. This replaces the change of 2026-11-16, which kept reader views on your Mac until you shared them.

Changed on 2026-11-16, and it stores less. A reader view made on your Mac is now kept on your Mac, with its comments, and nothing of it goes to Opyt's server unless you ask to share it. A shared view is a copy: comments made on it come back to your Mac. Reader views on the server no longer expire; each is kept until you take it down.

Changed on 2026-10-14, and it adds to what is stored. A reader view now holds your assistant's answer above the sources, so the answer goes to Opyt's server with the items. This version also spells out things that were already true: what a change from the web keeps of your Opyt's answer, libraries shared with you through the connector, what reader pages keep in your browser, and the Google Fonts both this site and reader pages load.

Changed on 2026-10-13, and it adds to what is stored. The connector can now change your library, and you can give an agent of your own a key. So the server now keeps, for a week, each change you ask for through the connector or an agent, and it keeps an agent's key as it keeps a reader's. Both are described below.

Opyt is operated from the United States. Questions: hello@useopyt.com.

Local install

On your machine, everything stays on your machine.

The Opyt you install is a local MCP server that reads and writes files in ~/.opyt on your own disk. It has no telemetry, no analytics endpoint, and no reporting call of any kind. Install counts come from PyPI and GitHub download statistics, which are public and need no code on your machine. What it imports comes from the accounts you connect and from public pages, and when it fetches a public web page it honours that site's robots.txt.

The one exception is deliberate and yours to trigger: if you connected your knowledge base to Google for the Opyt connector, or shared it (both below), your install uploads it, then uploads it again about once an hour while your Mac is awake, when it has changed (every five minutes for the first hour after you share or connect Google). What goes up is the export itself and the token saying which knowledge base it replaces — nothing else, and nothing about you. An install that has done neither holds no token, so this never runs at all.

The other is a reader view, also yours to ask for: the items it shows are uploaded once, as described below.

claude.ai and ChatGPT

What the Opyt connector stores.

If you connect your install to Google so that claude.ai, ChatGPT or another MCP client can read your library through the Opyt connector at api.useopyt.com, Opyt's server stores:

  • Your Google account ID, the number Google assigns your account, received when your install connects and each time a client signs in. It ties those sign-ins to your library. The sign-in asks Google for this ID only (the openid scope), so Opyt never receives your email address or name, and never sees your Google password.
  • A copy of your knowledge base — what your install imported: bookmarks, posts from people you follow, subscriptions, papers, pages, and who wrote them, which items you saved yourself, when each arrived, and which people you track (and when you confirmed them). Your install uploads it and keeps it current as your library changes; no connector tool writes to it. Its size and upload dates appear in the public totals (below), under a random key with no name attached.
  • Changes you ask for from the web — when you ask an assistant there to save a link, remove something, follow a topic or add a person. Your library lives on your computer, so the server holds the request until your own Opyt takes it, and then how it ended, so the assistant can tell you. What is kept is what you asked for (the link, the topic's lists, the item's id or the person's handle), when you asked, and what your Opyt answered, which can name the items or people it touched. Each is deleted a week after you asked, whatever became of it. Beside them the server keeps one timestamp per library: when your Opyt last asked for such requests, which tells it whether your computer is on.
  • Daily read counts — per day, which tool was used how many times, filed under a hash of your Google account ID rather than the ID itself. They show how often your copy is read, and feed the public totals below. Not what you searched for, and not which items you opened.
  • Sign-in records — each app that connects (such as Claude or ChatGPT) registers itself and receives tokens. The server keeps that registration (the app's name and callback address), the tokens, and the Google sign-in tokens behind them, encrypted on disk. Google's sign-in response, kept with those tokens, names only your account ID. Each call is checked with Google, and Google's answer is used in memory, not kept. Tokens work until they expire (at most a year) or are revoked; revoking one deletes the Google sign-in behind it at once and asks Google to revoke it. An expired record stops working but its encrypted file stays on the server's disk. The sign-in page sets cookies that last 15 minutes and tie your Allow Access click to your browser.
  • Libraries shared with you — if you paste someone's invite into the connector, the server keeps which Google account ID accepted it, the name its owner gave that library, which library it is, and a hash of the reader token the invite bought. The owner can see and cut off that reader like any other. It goes when they cut you off or delete the library; email hello@useopyt.com to have it removed sooner. While you connect Google from your Mac, the server also holds Google's one-time sign-in codes until you finish.

To receive those changes, an install that is connected to Google, or whose library an agent has a key to, keeps one request open to the server while your computer is awake. The request carries the library's own token and nothing else.

The connector never keeps the text of your searches (a query is matched in memory and never written down), which items you opened, or your IP address. Its log lines are the same as the sharing service's below: method, path and status code, with no query string, deleted after 30 days. How the connector works, tool by tool: connector documentation.

Your X and Substack sessions stay on your own computer. One exception: if Opyt runs on an assistant's computer that has no browser of its own, X sign-in happens in a private browser on Opyt's server. That browser runs on Opyt's server, so what you type into X's page passes through it; Opyt does not record it. The signed-in session is held in memory until your install collects it, at most ten minutes, then dropped, and that browser's files are deleted when it closes. Opyt never stores your password.

Opyt indexes what you import without any model, and sends your content to no model provider. Looking up papers and code asks OpenAlex, Semantic Scholar, arXiv and GitHub for their public records. Your AI client — where every question, answer, and all the reasoning happens — is governed by that client's own privacy policy, not this one.

Opyt does not read your knowledge base except as needed to operate the service or respond to abuse, does not analyze content across users, does not sell any of it, and does not use it to train models.

Reader view

A reader view is one private page on Opyt's server.

When you ask your assistant to open an answer's sources in reader view, the items it cited, in full, and the answer itself, the title and a line under it (often the question you asked) go to Opyt's server, which shows them on one private page. Nothing else of your library goes. Made from claude.ai or ChatGPT, the page is drawn from the connector's copy. Every few hours your Mac downloads a copy of all your reader views and their comments into ~/.opyt, so none is lost if the server is. Each paper's open PDF is fetched from where it is published, where the publisher's robots.txt allows it, and kept beside the page. The page's address is a long random id: anyone who has the link can read it, it is kept out of search engines, and nothing records who opens it. A page is kept until you take it down.

Anyone who has the link can also comment on the page. A comment is kept on Opyt's server with the name its writer typed, the passage it is about and when it was written, and everyone with the link sees it. Nobody signs in and nothing checks a name. The writer's browser remembers their name and can delete their comments; the server keeps no account and no address for them. Comments are deleted with the page.

A reader page keeps four things in your browser's local storage, read by nobody else: your theme, your chosen typeface, the name you comment under, and the keys that let you delete your own comments. It sets no cookies. It loads its typeface from Google Fonts, and shows a source's own images and video from the sites that host them, so those hosts see your browser's request; the page sends them no referrer.

Sharing service

If you share, the service counts — in daily totals.

The sharing service is reached only if you shared your own knowledge base, accepted an invitation to somebody else's, or opened sources in reader view. If you have done none of these, none of this applies to you. What it records, in full:

  • Access tokens as hashes — the SHA-256, never the token, so the database holds no usable credential. Alongside each: which one knowledge base it can reach, its role (the owner's, a reader's, or an agent's the owner gave a key to), the label its owner chose for it (the name you give your library when you share is shown to everyone you invite), and a random per-install id linked to no account, email, or machine. An invite is kept the same way, as a hash, with whether it is for a reader or an agent.
  • Storage accounting — how many bytes each shared knowledge base occupies, under a routing key assigned at random.
  • Daily read counts — per day (deliberately a day, not a timestamp), which reader used which tool how many times. Not which documents they read, and not what they searched for.

What is never collected: query text (searches are processed in memory and never written down), IP addresses (deliberately not logged, and a test fails if they come back), and per-document reading history. Read counts keep a day, never a time. A change you ask for from the web is not a read: it is kept with its time, for a week, as said above. Raw process logs are separate: each request leaves one line (method, path and status code, with the server's own timestamp, and no address or query), deleted after 30 days.

The service's totals are public at api.useopyt.com/stats, so whoever runs the service holds no more information than anyone else does. The full schema, checked against the code by an automated test, is in TELEMETRY.md — including its one honest limit: if exactly one person holds a grant, that knowledge base's daily total is that person's activity, and the service says so before their first query.

This website

No analytics, no cookies.

useopyt.com runs no analytics and sets no cookies. It loads its typefaces from Google Fonts, so Google sees that your browser asked for them. The one thing it keeps in your browser is your light-or-dark theme choice, in local storage, read by nobody. Invitation codes on /invite live in the URL fragment, which a browser never sends to a server — no server ever sees a code.

Your data, your call

Retention and deletion.

  • Everything on Opyt's server, yourself: say “unshare” in Opyt on your Mac and confirm. It deletes everything the server holds for your library: the copy, the Google link, every invite and every reader's and agent's key, the changes waiting for your computer, its reader-view pages and the daily read counts. The library on your Mac stays, and nothing is uploaded again unless you share or connect to Google again.
  • No longer have the Mac: email hello@useopyt.com and the same things are deleted for you.
  • Local install: delete ~/.opyt and everything on your machine is gone. That does not reach Opyt's server, so if you connected Google or shared, unshare first.
  • One item: the forget tool removes it from your library, and the server's copy drops it at its next upload, within about an hour while your Mac is awake.
  • One reader view: asking your assistant to take one down (reader_view with its link) deletes it at once, with its PDFs and its comments.
  • One reader or agent: unshare with their name cuts off that one alone; the copy stays up for everyone else.
  • A change you asked for from the web: its record is deleted a week after you asked.
  • One connected app: disconnecting or removing the Opyt connector in claude.ai, ChatGPT or another client invalidates that client's tokens. Removing Opyt's access in your Google account ends its sign-in too.
  • What lingers, and for how long: process logs are deleted after 30 days. The service database (the Google link, token hashes, read counts, reader-view pages and their comments, and pending web changes) is backed up nightly to seven rotating copies, so a deletion reaches the backups within a week; library copies are not backed up. Connector sign-in tokens that were never revoked stop working within a year, and after an unshare they can read nothing.

Opyt is not directed to children under 13, and we do not knowingly collect their data. If this policy changes, the new version is posted here; a change that expands what is collected will be called out, not slipped in.

Contact for anything on this page: hello@useopyt.com.

© 2026 Opyt Built with ❤️ in New Jersey Privacy Terms MIT