Security

Agents do the work. Policy decides what they may do.

Operkeel runs AI agents on your behalf, so the first security question is not only "is my data safe" but "what can the agents do, and who checks". This page answers both, and says plainly what we do not claim.

AI agent runtime governance

Every agent action is authorized before it runs.

Agents in Operkeel are governed at run time by Watchlight AI, a policy engine built for AI agents. It is not a prompt asking the model to behave: it is a check in code that the agent cannot skip.

Deny by default

Policies are written in Cedar, an open policy language. An agent may make a model call or a web search only if a policy permits it for that kind of work. Anything no policy permits is refused.

Checked twice, decided once

The agent checks each request inside its own process, and the gateway checks it again in a separate trusted process. Only the gateway's decision counts, so a confused or hijacked agent cannot approve itself.

Every decision is recorded

The decision to allow or refuse is written to your workspace's record before the action is carried out. If the record cannot be written, the action does not run.

Agents never hold your keys

Each run is its own operating-system process, started with an empty environment and ended when the work is done, at its time limit, or when you cancel. It has no API key and no database access. It can only ask the gateway.

Web content is treated as untrusted

What an agent reads on the web is screened for text that looks like instructions to an AI, and the agent is told to treat it as data. An agent cannot cite a page the gateway did not show it.

Hard limits on spend and steps

You set a monthly token budget; work pauses when it is reached. Each run also has a ceiling on reasoning steps and searches, so a loop cannot spend without end.

Actions that are not part of the service, such as sending email, publishing, changing a CRM, running code or driving a browser, are forbidden outright in policy, so they cannot be switched on by accident.

You stay in control of what ships

Agents research and draft. They do not publish or send anything. A reviewer agent and an editor agent check each draft against its sources, and a person approves it: PDF export is only for a version someone marked ready.

Your data is isolated from every other company

Each workspace is separated in the database by row-level security that is enforced by the database itself, not only by application code. A query that runs without a workspace in context returns nothing.

Your content is used to do the work you ask for. We do not use it to train AI models and we do not sell it. Model requests go to your model provider under your own account and its terms.

Keys and secrets

The model API key you connect is tested, then encrypted with AES-256-GCM and bound to your workspace, so a copy of the encrypted value cannot be opened for any other workspace. It is never shown again, never written to logs, and never handed to an agent. You can revoke it at any time, and runs in progress lose access at that moment.

Sign-in and access

  • Passwords are stored only as salted scrypt hashes.
  • Two-step sign-in with an authenticator app is available for every account, with one-time recovery codes.
  • Sessions are stored as hashes, so a copy of the database does not contain usable session tokens.
  • Team members have roles, and every command checks the caller's permission.
  • Sign-in attempts and public forms are rate limited.

Chat clients (MCP)

When you connect Claude, ChatGPT or another MCP client, it signs in through OAuth 2.1 with PKCE, and you approve the connection in Operkeel. Access tokens last one hour and are tied to this service; refresh tokens rotate, and reusing an old one revokes that client. Each call re-checks your membership, your permission and your subscription, is rate limited per workspace, and is recorded with the client that made it. Actions that need a person at the desk, such as connecting a key, are refused from chat clients. You can disconnect a client in Settings.

Payments

Subscriptions are handled by Stripe. Card details go to Stripe; we never see or store your full card number.

Records you can read

Your workspace keeps a record of key changes, commands that change something, and every allow-or-refuse decision made for an agent, so you can see what happened and when.

What we do not claim

We would rather be exact than impressive. Operkeel does not yet hold a SOC 2 or ISO 27001 certification, and we have not had an independent penetration test. Hosting is on Railway, whose infrastructure security is described in its own documentation. If your review needs something specific, ask us through the support page.

Reporting a problem

If you believe you have found a security issue, tell us through the support page and we will reply to you directly. See also our Privacy policy and Terms of Service.

Ask your market a real question.

Create a workspace, start your subscription, and get a sourced answer today.

Sign up Try the live demo