Trust center

What protects your data, and where it is enforced.

Each line on this page is a mechanism in the code rather than a promise about it. A security review can start here and find every one of them.

Report a vulnerability

Found something that looks unsafe? Tell us before telling anyone else, and we will work with you on it.

Accounts and access

Passwords are hashed with scrypt
Where this server keeps a password itself, it stores only a salted scrypt hash, with the cost written beside it. The password is never stored or logged.
Approvals need a recent password
Deciding a gate, going live and releasing the emergency stop ask for proof that you signed in recently, not just a session that is still open.
Session cookies that script cannot read
The session cookie is HttpOnly and SameSite, and Secure over HTTPS, so script on a page cannot read it and another site cannot send it.
Throttled public endpoints
Sign-in and the other endpoints anyone can reach are rate limited per address and per account, and can require a bot check when one is configured.

Each workspace is isolated in the database
Every table holding a workspace's data has Postgres row-level security forced, scoped to that workspace. The server refuses to start on a connection that could bypass it.
Saved credentials are encrypted
Integration keys and other secrets are encrypted with AES-256-GCM under a key held outside the database, and are shown back only as a fingerprint.
Secrets are redacted from transcripts
Secret patterns are removed from everything an agent run stores or streams, so a key pasted into a ticket does not end up in a log.
Deleting a workspace deletes its data
After a grace period in which it can be called off, every row a workspace owns goes with it, and its stored files are erased by a sweeper that keeps a ledger and retries until they are gone.
Your data can leave with you
A workspace can be exported as a file holding its work, its decisions and their history, and no credential, token or fingerprint of one.

A person approves every production write
Merging and deploying need a single-use token that is minted only when a person approves the gate. The database refuses to change a token once it has been spent.
One switch stops every agent
An admin can stop every agent on a project or a whole workspace at once. Starting them again takes an owner and their password.
A tamper-evident audit log
Audited actions are written to an append-only log in which every entry's hash covers the one before it, so an entry changed, removed or inserted later breaks the chain where it happened.

A content security policy with a fresh nonce
Every page is served with a content security policy whose script nonce is minted for that one response, and with HSTS over HTTPS.

Documents

Each legal page is a template until whoever operates this deployment completes it, and says so at the top.

Every integration is mocked until you add credentials, so you can watch the whole organization work on your product before it touches a real system.