Trust, andhow to leave.

The engine that runs your agents is closed source, and you use it through our hosted API. You can't run it on your own servers.

The code that talks to it is open under the MIT licence, in our public repo (opens in a new tab): the TypeScript client (@argentic-labs/client on npm), the Python client, the plugin for Claude Code and Codex, the examples and templates, and the crash-test tool. You can read exactly what your app sends us.

About the engine itself we publish results, such as the studies and benchmark numbers on Research, and not how it is built.

You don't have to take the engine's word for what it recorded. The journal file carries a SHA-256 hash chain that a short script, run by you with nothing but Node, recomputes; Proof of what ran has it. Keep the final hash somewhere of your own and any later copy can be checked against it.

Each agent's journal downloads as one JSON file. It holds every message, model turn, action and tool result, what the agent chose to remember, and the hash chain over all of it. Use Download journal on any run page in the dashboard, or call the API with your key:

• GET /v1/agents/{agent}/journal/bundle returns the whole journal. A file is capped at 4 MB; past that, pass to_seq to take it in parts, or page through GET /v1/agents/{agent}/journal.
• GET /v1/agents/{agent}/memory lists the agent's decisions, facts and plans (kind=decisions, facts or plans). They are built from the journal, so the file has them too.

To take everything at once, owners and admins can press Export everything in Settings. You get one zip with a journal file for each agent, and a note naming anything it left out, such as a journal over 4 MB.

No other engine can load these files and carry on a run. They are plain, documented JSON that any tool can read, and they are your record of what happened.

To leave for good, press Email us to delete at the foot of your account page or your organization's Settings, or write to hi@getargentic.com. We delete the account's memory and journals by hand and revoke its keys. Stripe keeps billing records for as long as the law requires.

If we ever close the service, the terms give you at least 30 days' notice, which is time to download everything.

You bring your own model provider key and send it with each call. We use it for that call, to the provider you named, and we never write it to the database or to our logs. It is removed from error messages, and our request logs keep the method, path, status, account and timing, never headers or bodies. Your provider bills you directly for the model.

The key does pass through our servers on every call, because the engine makes the model call for you. Give Argentic a key of its own with a spend limit set at your provider. If anything looks wrong, you can revoke that one key and nothing else breaks.

Your agents' data is stored in Postgres at Neon, on AWS in us-east-1 (Northern Virginia, United States). Neon also runs the API. Your data is encrypted at rest with AES-256 and reached only over TLS. Vercel serves this site and the dashboard. We are based in Australia, so your data is processed overseas.

Neon can restore the database to any moment in the last 6 hours. A problem found later than that may not be recoverable, so keep your own copy of any journal that matters to you.

The full list of companies that handle your data, and what each receives, is under Subprocessors.

• Memory and journals stay while your account is active.
• Cold storage for journals is built and off until we switch it on. Once it is on, a journal idle for a couple of days is compressed, nothing is dropped, and the next read or wake restores it exactly.
• The stored reply to a finished call, which a retry with the same key gets back, is kept for 30 days. The journal still has what the agent said.
• During the beta we may reset data, and we will warn you first when we can.

The privacy policy has the details, including analytics and the waitlist.

There is no public status page yet, and the terms make no uptime promise. We plan to add a status page after launch.

An automated check calls the site and the API a few times a day and alerts us when either goes down or comes back. Argentic is run from Sydney (UTC+11) and nobody is on call around the clock, so a fault overnight there can wait until morning. If something looks broken, ask on Discord (opens in a new tab) or email us.

Changes merged on 21 and 22 October had written internal security reviews, and later audits covered sign-in, access control, headers and the firewall. The reviews cover who can reach which data, isolation between organizations, what analytics might capture, outbound requests from webhooks and tools, injection, secrets in code and the browser's security headers. Anything serious is fixed before it ships or before the feature it is in is switched on. Ten review rounds, on 21 and 22 October, found no critical or high-severity issue.

We run automated tests before every release that check one organization can't read another's runs, memory, tools or approvals. When someone at Argentic opens your organization's data, owners and admins can see who it was and the reason they gave, in Settings.

What we haven't done: nobody outside Argentic has audited us, there has been no penetration test, and we have no SOC 2 report or bug bounty. If you find a problem, the security page says how to report it.