Security & privacy

Know what your agent sends to Pushary.

Pushary is built on data minimization. Approval requests can include a command or diff, and task titles can include part of a prompt. Pushary-launched transcripts sync only after local redaction and encryption for the owner's phone and documented compliance recovery. Here is exactly what that means.

Approval data

The approval question, tool name, and notification text
The command or diff included in an approval request, with secrets redacted by the hook
Your decision and the information recorded in the audit trail

Session data

For Claude Code, the first 120 characters of a prompt can become the session task title; task titles can be disabled
Sessions launched by Pushary sync transcripts by default after local redaction and encryption
Transcript keys are wrapped for the owner's phone and documented compliance recovery; this is not zero-knowledge encryption

How a decision flows

The hook intercepts

A lightweight hook runs in your agent (Claude Code, Codex, Cursor, Hermes, or any MCP client) when a tool call needs a decision. The agent continues running on your machine; the approval request can include a command or diff.

Transcript content is encrypted

Pushary forwards the approval request to its API, applies your per-tool policy, and sends a push when a human decision is needed. Synced transcripts use a separate path that redacts and encrypts content on the agent machine for the owner's phone and documented compliance recovery.

You answer in place

You approve, deny, or answer from your lock screen. Decision page links are signed. Treat them as sensitive and share them only with the intended reviewer.

The agent continues

Your answer returns to the hook and the agent keeps working. The request and decision are recorded in the audit trail. Approval text can contain a command or diff.

Honest about encryption

Pushary has two data paths. The decision path carries the question, tool name, and approval context, which may include a command or diff, preserving a server-side, queryable audit trail. Pushary-launched transcript sync is on by default, but redaction and encryption happen on your machine first. The session key is wrapped independently for the owner's phone and Pushary's compliance recovery key; authorized admins can recover one exact session for a documented legal obligation. You can disable sync on the agent machine. Traffic is over HTTPS, decision links are HMAC-signed, and pending questions expire from cache on a short timer.

See the data flow and disable task titles or transcript sync.

Frequently asked questions

Does Pushary see my code?

Approval requests can include commands or diffs, with secrets redacted by the hook. Claude Code task titles can include the first 120 characters of a prompt. Sessions launched by Pushary sync transcripts after local redaction and encryption for the owner's phone and documented compliance recovery. Task titles and transcript sync can be disabled separately.

Is Pushary end-to-end encrypted?

No. Pushary uses provider-managed encryption for synced transcripts, not end-to-end or zero-knowledge encryption. Redaction and encryption happen on the agent machine first. The session key is wrapped independently for the owner's phone and Pushary's compliance recovery key, so authorized admins can recover an exact session for a documented legal obligation. The decision path remains minimized and queryable for the audit trail. You can disable transcript sync on the agent machine. Traffic is over HTTPS.

What does Pushary store, and for how long?

Pending questions are held briefly in Redis with a short time-to-live and then expire. The audit trail records each question, the tool involved, and the human decision (who decided, when, and under which policy), so you and your team have a record. Approval text may include a command or diff. Encrypted transcripts are a separate data path.

How should I handle decision links?

Treat a decision link as sensitive: it is intended for the reviewer receiving the request. Signed links protect against tampering; they are not a reason to publish or forward a live approval link.