Cealt is being built to check every agent action routed through it against your organisation's own rules. Clean and minor actions go through; for minor issues the agent gets a suggested fix. Critical ones wait for the person your rules name, forbidden ones are refused, and every decision is recorded.
Actions that break no rule go through. Minor issues go through too, recorded, with a suggested fix for the agent’s owner.
When an action breaks a critical rule, it waits for the person your document names — not a shared queue.
Designed to give a quality score per agent and action type, with concrete improvement suggestions for the people who run them.
In the prototype, on synthetic data. Platform hooks and target-system logs: design goal.
Cealt's coverage report shows how many of your registered agents sit in each layer, so you know what is under control and what is not.
Design goal: for Microsoft 365 and Google Workspace, the agents no longer hold the tool keys. Cealt's key broker uses them, and you can revoke its access at any time. Those systems' own logs are matched against Cealt's record to flag actions that went around it, including from agents nobody registered.
In organisations with a security platform, Cealt works next to it. In smaller companies, Cealt may be the only control in front of the agent tool calls routed through it, so we are explicit about what it does not do: it does not inspect network or device traffic, it does not detect prompt injection, it does not provide model security, and it does not classify content it has not been told to look for.
For banks, holdings and public bodies. The same package runs in your network, with your own key vault and identity provider, and it does not call home.
For companies that would rather not run it themselves. Sign in with Microsoft or Google. Each customer's data is kept apart. A daily fingerprint of your record goes to a place you control, so you can check that no one, including us, has rewritten what was recorded up to that day.
The hosted option opens once Cealt's company is formed, data processing terms are in place and counsel has reviewed them.
Whether an action goes through or waits is decided by approved rules, never by a model's judgement.
Cealt is designed so that a rule or setting takes effect only when two people other than its author sign it, and the change is recorded.
Each decision points to the clause and the people behind it.
We price per organisation per year, by employee band. Your bill should never be a reason to leave an agent out.
Hosted by Cealt. For companies that would rather not run it themselves.
In your environment. For banks, holdings, regulated organisations and larger companies.
Cealt is not for sale yet. Prices are on request. Every feature above is in design or in a prototype on synthetic data.
We work with a small number of organisations at a time. Write to us and we will reply personally.
Tell us your name, organisation and role. If you have an invite code, mention it.
Or copy the address: hello@cealt.ai
There is no form on this site. We use your email only to reply to you. Privacy notice
No. Cealt is in private development. Design partners help shape the first version, and nothing is in production yet.
That is what Cealt is designed to avoid. Actions that break no rule and minor issues pass without waiting for a person; only actions that break a critical rule wait. Your organisation decides what counts as critical and what counts as minor; two people approve each such rule, and minor issues are recorded too.
No. Those systems keep their own approvals. Cealt is designed to sit in front of AI agents' tool calls and send the actions that need a person to the person your documents name.
With on-premise deployment, no: documents, rules and the decision record stay in your network. With the hosted option, Cealt keeps them, separated per customer; where they are hosted will be stated before the option opens. Either way, your agents' own AI models still see what those agents send and receive. Cealt does not change that.
Your Cealt owner writes each rule in the rule editor and tests it in a simulation; two other people sign it before it takes effect. Drafting rules from your existing documents is being designed for the Enterprise package, using a model inside your network; a person reviews every draft.
A named owner, for example in internal control or financial control. They write and test rules, keep the agent catalogue current and follow up on the coverage report. Cealt is designed so that they cannot put a rule into effect alone: two other people sign it.
It is not a security platform. Where you have one, Cealt works next to it. Cealt checks the agent actions routed through it against your own rules, stops only what those rules mark critical, refuses what they forbid, and shows which of your registered agents it cannot see.
Per organisation, per year, by employee band. Agents, actions and approvers are unlimited, so registering every agent never raises your bill. Prices are on request; Cealt is not for sale yet.
A sample authority document (anonymised is fine), a few working sessions with your internal control team, and honest feedback.