Privacy summary
Version 2026-11-13. Part of the Terms of Use.
Operator. Sipki LLC, New York, USA. Contact: support@sipkigroup.com.
What we store
- Account: random id (
dw:…) and the display name you type. No e-mail, phone, password or payment data. When you make the account's first passkey: which version of the Terms you agreed to and when (kept after Delete account, with the opaque id, as the record of the agreement). - Passkeys: public key, credential id, signature counter, device label. Private keys and biometrics stay on your device.
- Sessions: a SHA-256 hash of the cookie; expire after 7 days.
- Agents: name; API key as a peppered hash plus short prefix.
- Rules you sign: mandates, policies, allowances. "Closed data" values are stored as typed.
- Agent actions: tool, class, target in clear (e-mail address, file path, host), amount, decision, time, argument hash. Masked previews (≤300 characters, ≤4 KB text) shrink to ~80 characters once decided and to the hash after 30 days.
- What you signed: when you approve a request, the exact object the approval page showed you is kept with your signature: the action, both masked previews, the masked text of the written and changed files it showed, and the command names of the line. It is kept 30 days from signing, so the evidence pack can prove what you saw (it is in the pack), then deleted; a pack made later says that approval's view can no longer be checked. Delete account erases it.
- Claude Code hook: sends each tool input as given (full command line, paths, written file text); it is masked on the server before anything is written to disk. Kept until you delete the account, besides the action itself: paths and hashes of executed project files and the hosts they name, your project folders, every path a write touched, the command names of each command line, the project and outside paths a command uses, and the Claude Code session id of each decision. When the agent changes a file a later command executes and the witness could not read the change, the hook sends that file's text, masked, up to 4 KB, for the approval page only; it is cut with the other previews once you decide (the copy you signed is kept 30 days, see above). When a write looks risky (it talks to the network, builds the code it runs, or its content is not visible: a copy, a download), its masked text (≤4 KB) is kept so that the page of a later run of that file can show it; it is deleted once you decide that run, or 24 hours after the write (checked every hour, whether or not the agent runs again). With each call it also sends the working directory, the project directory, your home path, your temp folder path (
$TMPDIR), the Claude Code session id, the git branch of the project, and its answers to the witness's file questions (does a path exist, where it resolves; for a hard link of a credential file, the name of that file; for an archive the call extracts, the link entries inside it; for a git branch switch, which files it changes; for a file written outside the project, the path, hash and size of the copy kept on your computer); for Bash also: where the Python tools named in the command resolve on your PATH, a SHA-256 hash of your gituser.emailwhen the command line itself shows that address, and a summary of your Claude Code sandbox settings (platform, allowed and denied paths, allowed domains, excluded commands, local sockets and macOS service names, permission mode, the names, not the contents, of.envfiles in the project folder, the names, not the values, of environment variables that look like tokens, and which of your config files hold a credential, by file and kind, never the credential). Requests carry your agent key. The full list, field by field: docs/data.md. When the witness asks about the build and test files a command may run, it sends their paths, SHA-256 hashes and sizes, and the values of git settings that start programs (such ascredential.helper,core.sshCommand, aliases); before they leave your machine the hook masks the secrets inside each value (credentials in a URL, well-known token prefixes, the word afterpassword=,token=,Authorization:,Bearer), and keeps the program path and flags; a secret written in another form could stay, so keep secrets out of these settings. Your prompts, Claude's replies and tool outputs are not sent. - MCP proxy: the full arguments of each call (stored masked, as above); once per run the name, description and input schema of every tool of your MCP servers, and a coverage report that names MCP servers not behind the proxy. Kept until you delete the account.
- Optional: push subscriptions; Telegram/Slack/e-mail channels (secret encrypted, AES-256-GCM); Team form (e-mail and text, deleted after 180 days).
- Quick start accounts (
deus-witness try): the same as an account above, plus the expiry time, a SHA-256 hash of the one-time claim link (deleted at your first passkey touch) and the drafted mandate form of your project (project and temp paths, command and domain lists). No passkey, no name and no journal exist before your first touch. - Counters: daily totals; which onboarding steps an account passed and when (the day, to count "came back on another day"); whether an account is ours (internal).
Which accounts are erased by time
A quick start account nobody has claimed with a passkey, and everything in it, is erased 7 days after deus-witness try made it if its hook never called the witness (every row is deleted; there is no journal yet) and its agent key stops working; while its hook calls, it stays until 90 days after the last call. Free for one person. If this ever changes, existing accounts stay free on the limits in force at that time, and we announce it at least 30 days in advance. Free limit: 10000 actions per day per account and 4000 per agent (reading the project's files does not count; a quick start account before its passkey: 2000). An account in use is kept: it is erased only after 90 days with no call from your hook and no sign-in, and 14 days before that we send a push to your devices that have notifications on. The erase of an unused account with a passkey is the same as Delete account below (the journal becomes its hash skeleton); a team, or a person who signs for another team, is never erased for being unused. Backups follow the 15-day rule below. Honest limit: until your first touch, whoever has the claim link can make the account theirs; the link goes only to your browser (in the part after #, which browsers do not send to any server), and it is not in the command's output or a file: with a browser it is opened there (the browser keeps it in its history until you claim the account), without one it is shown on your terminal only (written to /dev/tty).
What we do not store
IP addresses in the database (the rate limit keeps them in memory only and drops each at most 2 hours after its last request; they do appear in the reverse-proxy access logs, see Retention); request bodies in logs; third-party analytics or ad cookies. Unmasked text is never written to disk: to check risk, the server holds the unmasked text of the last 64 calls in memory only; newer calls push it out and a restart clears it.
Append-only journal and deletion
Each decision judged against your mandate enters a hash-chained per-account journal (a few refusals stay in the operator's database only, see the concepts page); checkpoint hashes are anchored in Bitcoin via OpenTimestamps and time-stamped by FreeTSA. Delete account erases the data above (MCP tool lists and coverage reports included) and replaces your journal with a skeleton: a signed final checkpoint plus the list of entry hashes, so other accounts' proofs and packs you already downloaded stay verifiable. Action targets, amounts and arguments are gone from the server. What remains: entry hashes, checkpoints, the key registry and the opaque account id. Team policies (org_policies) are not deleted with the account either: each is bound to an entry of the registry log, which is shared and append-only. Packs you exported before deletion are yours and cannot be recalled.
Hosting and backups
Hostinger VPS, Boston, USA. Daily backups, each generation encrypted with its own key; a generation and its key are destroyed after at most 15 days. Copies on the operator's computer (same encryption) and in Microsoft Azure storage (encrypted, deleted after 15 days). So deleted data is unrecoverable from backups after at most 15 days.
Third parties (only if enabled by you or the operator)
- Browser push services (Google, Mozilla, Apple, Microsoft): opaque ids only.
- Telegram, Slack, operator's e-mail relay: "an agent is waiting" plus a link.
- Microsoft Azure (backup storage, USA): encrypted archives only.
- OpenTimestamps calendars, FreeTSA (freetsa.org, a time-stamping authority), mempool.space, blockstream.info: hashes only.
Retention
While the account exists and is used (see above: 90 days after the last call of your hook or your last sign-in); an unclaimed quick start account its hook never called, at most 7 days. Sign-in challenges: 5 minutes. Push subscriptions: until gone or 5 failures. Application logs: ~50 MB rolling. Reverse-proxy access logs (these include the IP address of a request): up to 30 days, not in the database. Backups: at most 15 days. Anchored hashes: permanent.
Your rights and requests
Your rights. You can see and export your data: “Download evidence pack” exports your journal and receipts, and the objects you signed in the last 30 days (not every table). You can delete your account in the dashboard at any time (what remains is said above). For a copy of anything else, a correction, or any other request, write to support@sipkigroup.com; we answer within 30 days. If you are in the EU or the UK, we process your data to provide the Service you asked for (contract) and to keep it secure (legitimate interest), and you may complain to your data protection authority.
We do not sell or share personal information, and we do not use it for advertising or to train AI models.
Other people's data. Your agent's actions may contain other people's data (for example an e-mail address it writes to). You are responsible for having the right to send it to us; it is masked and kept as described above.
Children. The Service is not for people under 18; we do not knowingly collect data of children under 13.
Security incidents. If a breach affects your data, we will tell you without undue delay and as the law requires.
Changes. We post a new version of this summary here with its date, and announce material changes at least 30 days in advance.