Control the intelligence running your organisation.
One place for your people to work with AI — and one control plane to discover, govern, secure and account for every model, agent, identity, data source, tool and action.
AI did not enter the organisation through one door.
It arrived through hundreds.
You cannot control what you cannot see.
Imperium discovers the AI estate. A scanning wave passes across the network. Unknown nodes become identified — and every one receives an owner, a department, a location, a trust level and a policy.
| OBJECT | KIND | DEPT | TRUST | POLICY |
|---|---|---|---|---|
| claude-enterprise | MODEL | ENG | MED | governed |
| llama-70b-frankfurt | MODEL | AI-R | HIGH | local-only |
| finance-agent-07 | AGENT | FIN | MED | approval |
| github-mcp | MCP | ENG | LOW | approval |
| copilot-ws-0412 | CLIENT | ENG | MED | repo-scoped |
| erp-summariser | AUTOMATION | OPS | HIGH | governed |
| internal-file-mcp | MCP | IT | HIGH | governed |
| chatgpt-browser-×214 | CLIENT | ALL | UNK | unmanaged |
One control plane.
Every AI.
Every model, agent and tool call becomes a governed request: who is asking, what it may touch, where it may run, what it may do afterwards — with evidence.
One place to work with AI.
Your AI doesn't need to live in five different products — each with its own conversation, context, history, credential and model.
A workplace for people and agents. A control plane for the organisation.
OTOPIA brings conversations, tasks, shared context, agents and human decisions into one working environment. Imperium governs the authority and actions behind that work. Together they make agent collaboration practical and accountable.
Capabilities are enabled according to deployment and licensing. Imperium does not require Observatory, and Sentinel, Observatory and OTOPIA are distinct parts of the ecosystem rather than one product.
Your work should not disappear when the window closes.
One session, carried by Imperium — not by the tool you happened to open it in.
THE MODEL MAY CHANGE. THE SESSION DOESN'T.
AI doesn't only read prompts.
Documents, images, conversations and messages enter the same governance model.
CONVERSATIONS ARE AI INPUT TOO.
A representative request, moving through Imperium.
Understand the request, apply the organisation’s controls, execute inside that boundary, check the consequential action, and keep the evidence.
ACTION BLOCKED
The summary drew on an unverified external source, and this organisation does not allow confidential material of that provenance to leave the company. The attempt, the reason and the evidence are preserved.
Nothing privileged executes until the gate says ALLOW.
Before an agent takes a consequential action, Imperium checks whether the actor has the authority to perform it in the current context. The action proceeds, stops, or waits for human approval — and the decision and its reason are recorded.
Illustrative scenarios with synthetic data. Pick an action above to see how the outcome changes with the actor, the authority and the context.
Not a visualisation of what a control plane might become — an operational platform. The screens below are the product interface, shown with synthetic demonstration data.
One view of the AI estate.
Models, agents, providers, usage, security, policy and cost — in one operational view.
See what every AI is doing.
Request to model, tool and final action — identity, policy decision, model, locality, provider, cost — recorded on the wire, never inferred.
AI does not decide what it is allowed to do.
Policy does.
Set organisation-wide boundaries, then delegate appropriate controls to teams and applications. Local flexibility remains within the limits the organisation established. Try the inheritance model in the interactive demo →
Policy that follows the organisation.
Global controls with department, team and user inheritance and controlled overrides — tenant → org → department → unit → project → group → agent.
AI actions have a chain of authority.
AUTHORITY SHOULD NARROW AS AUTONOMY EXPANDS.
LEAST PRIVILEGE. FOR AI.
When work is delegated, Imperium helps organisations see who authorised it, what the agent could access, and which actions required approval. Access granted for a task is time-limited and can be withdrawn.
Don't just know that something happened. Know how.
The AI flight recorder reconstructs an execution: who called whom, who delegated what, which data moved, how trust changed, where policy evaluated, what it cost, where the blast radius grew — at any point in time.
Control every agent.
Discover where agents run, what they do, what they cost and whether they remain inside policy.
Policy defines what is permitted. Imperium selects an appropriate permitted path.
DECISIONS NEED CONTEXT, NOT JUST RULES.
A decision is not the end.
Governance continues after it — through a workflow that survives waiting, approval, failure and rollback.
- COLLECT STATE
- CHECK DEPENDENCIES
- REQUEST APPROVAL
- WAIT
- APPROVED
- RESTART SERVICE
- VERIFY HEALTH
- ROLLBACK IF FAILED
- CAPTURE EVIDENCE
GOVERNANCE CONTINUES AFTER THE DECISION.
AI SOLVES IT ONCE. AUTOMATION REUSES IT. A successful resolution can be reviewed, approved and reused as a governed workflow — a person decides whether it earns that, not the system.
Who watches the control plane?
Imperium Sentinel provides an independent integrity layer for the control plane. It helps detect unauthorised changes and preserves security evidence even when the primary platform is disrupted.
Know what left the organisation.
Every execution carries its destination — in-house, private or external — recorded by the appliance, never guessed from a URL. Breaches and over-permissioning are on the same map.
One control plane. Every layer of AI.
Nine capability domains — a way to evaluate organisational AI control, not a processing order.
01DISCOVER
Continuously find models, agents, clients, tools and MCP across the organisation.
02GOVERN
Apply identity, organisational policy and inherited controls.
03DECIDE
Resolve allowed outcomes under policy, risk, cost and operational constraints.
04CONTROL
Broker credentials, routes, models, tools and privileged authority.
05OBSERVE
Trace requests, agents, models, tools and runtime behaviour.
06VERIFY
Propagate provenance and trust and detect behavioural change.
07OPTIMIZE
Control tokens, cost, models, budgets, GPU and compute.
08PROTECT
Enforce sensitive boundaries before actions execute.
09EXECUTE
Run durable governed workflows across agents, tools and systems.
Run it where your data lives.
Principles for organisational agent work.
How we think this ought to work — offered as a way to evaluate any organisational AI control system, including ours.
- Attributable identity. Every human and agent action has an identity behind it.
- Bounded delegation. Delegated authority has a defined scope and can be withdrawn.
- Checked before execution. Sensitive actions are checked before they happen, not reported afterwards.
- Context travels with information. Information retains its origin and classification.
- Human oversight. People can inspect, approve, interrupt and review consequential work.
- Reconstructable. An organisation can follow the chain from intent to outcome.
- Open collaboration. Work remains possible across approved models and tools.
- Impact is traceable. Operational effects can be linked back to the agent activity that caused them.
A Tracston working model, published so the vocabulary and the evaluation criteria are in the open. Not a standard, and not a claim of adoption beyond our own products.
IMPERIUMWork with AI.
Without losing control.
One workspace for people. One history across models and agents.
One policy for the organisation. One evidence trail for every important action.
FROM HUMAN INTENT · TO AI ACTION · TO REAL-WORLD IMPACT.