Privacy

Svanetra keeps how you like your AI tools to work, as short rules such as “Give the root cause first, then the fix.” It never keeps your code, your files or your chats: when a tool reports a turn so Svanetra can learn from it, the text is read once and deleted, and only the preference learned from it is kept. Your graph is private until you share it, and you can delete everything at any time.

• Your account: your email address, for signing in and for friend invites. It is never shown to other people. Also your handle, your display name, and a picture link if your sign-in provider gives one.
• Your taste graph: the rules your tools saved or you wrote, their domains and links, how often each was used, and a history of changes so you can undo them.
• The evidence behind each rule: what kind of signal it was (something you said, a correction, an approving reaction, a thumb), when, from which tool, and how much it counts. No text.
• What was served: which of your rules each tool was given and how they were scored, with a list of numbers computed from the task summary (an embedding) to match tasks to rules, never the task's text. Kept for 180 days.
• What each reported turn was: a correction, a stated preference, praise, or none of these, and which rule it was about. No text. Kept for 180 days.
• Kinds of work, named from a broad area and a technology, such as “Debugging · Go”, with an embedding at their center.
• Preferences you give for one chat only (“for now”), for up to 12 hours.
• Your answers to the weekly question and the taste check, and your thumbs on rules.
• Connections: which AI tools you connected, and when each last synced or was used. For hook tokens, only a fingerprint (a SHA-256 hash) of each token, and when it was last used.
• Daily counts for each tool: chats served, corrections and saves, for the Impact screen.
• Sharing: what you share and with whom, email addresses you invite until they join, and your stars, follows and reports.

Your code, file paths, chat history or messages. When a tool reports a turn, through its hooks or when its agent calls note_feedback, it sends your message, up to five of your earlier messages in that chat, and the end of the agent’s last answer. Svanetra holds them only until it has read them, usually within a minute, then deletes them; chat ids are kept only as a fingerprint. A rule or note that contains code, a file path, a link, an email address, a username, a phone number or anything that looks like a key or password is refused, not stored. So is a rule that tries to give a tool permission to act, rather than describe a style.

To give your rules to the tools you connected, to show your graph so you can change it, to answer your questions in the graph chat, to tidy duplicate rules overnight, to measure whether your rules work, and to share what you choose to share. Svanetra also tunes its own settings each week from everyone’s outcomes taken together, as counts and averages; how many people hold a preference is used only when at least five do. In about one session in fourteen it holds back the one rule that adds least, to measure what your rules prevent; you can turn this off in Settings. No ads, no selling data, and your rules are not used to train AI models.

WorkOS

Sign-in. Sees your email address, name and how you signed in.

Neon

Our database. Stores everything listed above.

Railway

Runs our servers.

Vercel

Runs this website. Requests pass through it on their way to our servers.

OpenAI

Answers the graph chat, tidies rules overnight, helps check saves, turns rules and task summaries into numbers so the right ones match your task, and reads each turn your tools report, once, to find the lasting preference in it. It sees the rules involved, your question, and the turn’s text. OpenAI doesn’t use data sent through its API to train its models.

TypeSafe

Its Jev model checks each save (whether it repeats, refines or conflicts with a rule you already have), sorts each reported turn (a correction, a preference, praise, or none of these), and, for a sample of turns, judges whether the rules served were followed. It sees that turn’s text once and the rules involved.

Google, GitHub

Only if you sign in with them. They tell us your name and email address.

Nothing, until you share. Sharing is friends-only unless you choose a link or public. People you share with see the shared rules and their links, your handle and your display name, never your email address. A public graph can be seen by anyone and is listed in Community.

A cookie keeps you signed in. Your browser also remembers your theme and layout, and your name, handle and email address so they show quickly. No analytics or advertising cookies. If you install Svanetra’s hooks, a file on your computer keeps the last five messages and the end of the last answer of each chat for a day, so the next turn can be understood; it leaves your computer only as described above.

• See and edit every rule, undo changes, remove a connected tool, revoke a hook token and stop sharing, at any time.
• Turn holdouts off in Settings, and see why Svanetra believes each rule in the rule panel.
• Delete your account in Settings. That removes the account and everything in it right away. Copies in our database host’s backups expire on their own within a few days.
• Our servers keep short-lived logs, such as request times, errors and IP addresses for rate limits, to fix problems.

Svanetra isn’t meant for anyone under 13, or under the minimum age for online services where you live.

When this page changes, the date above changes too; for big changes we will tell you in the app or by email. Questions or requests: hello@svanetra.com.