Products / AI agent control engine

Mandate

AI agents do the work. Companies set the rules. Mandate gives every agent an identity, a defined scope and a complete audit trail.

Request early access

The problem

AI agents now reach live databases, APIs and internal tools the way an employee would. Increasingly they act: updating records, approving changes, moving money.

Most run on static keys with broad grants, because fine-grained scoping is hard to build in-house. Some route credentials through a gateway the enterprise does not control. When something goes wrong, nobody can say exactly what the agent did, or on whose authority.

How it works

  1. Give each agent an identity

    Each agent is declared as a spec and registered with an identity, a version, an owner and a named business capability.

  2. Grant only what the role needs

    Access is scoped to the tools and data that capability requires, with short-lived credentials issued per task instead of standing keys.

  3. Check every action before it runs

    Each call is allowed, denied with a stated reason, or held for a person to approve.

  4. Record everything

    Every access, finding, attempt and decision is logged, so when someone asks what an agent did, there is a precise answer.

Treat an agent like a new hire

A new employee gets a badge, a role and access to what the job needs, and their work can be reviewed. Mandate gives an agent the same.

A new employee gets An agent on Mandate gets
An ID badge A verifiable identity, listed in your directory
A defined role A named business capability
Access to the information the job needs Least-privilege access to the tools and data that capability needs
Permission to take certain actions Action boundaries, checked before every call
Work that can be reviewed A complete audit trail

Watch it work

A procurement agent is reviewing an invoice. It is allowed to read orders and supplier history. It is not allowed to change a supplier’s status.

A procurement agent governed by Mandate Step 1, allowed: the agent reads the invoice and matches it to the purchase order. Step 2, allowed: it finds six orders from one requester, each just under the approval limit. Step 3, denied: it tries to put the supplier under review, which is outside its capability. Step 4, recorded: every read, finding, attempt and decision is logged with the reason. A person makes the final call. STEP 1ALLOWED Reads the invoice Matches the purchase order from a regular supplier. STEP 2ALLOWED Finds a pattern Six orders from one requester, each under the approval limit. STEP 3DENIED Tries to act Wants to put the supplier under review. Not its capability. STEP 4RECORDED Records it all Every read, finding, attempt and decision, with the reason. A person makes the final call. Mandate makes sure the agent never made it first.
The agent can flag the pattern. Only a person can act on the supplier.

Where it runs

  • In your cloud. Mandate deploys inside your own environment. Credentials and audit records never leave it, and logs go to your own logging stack.
  • In your directory. Each agent’s identity can appear in Active Directory or Microsoft Entra ID next to your people, so access reviews, groups and offboarding cover agents too.
  • With the agents you build. Engineer-built agents on LangChain, CrewAI or custom code, calling your APIs, databases and internal tools.

Who it is for

  • Platform and security teams running engineer-built agents on frameworks such as LangChain, CrewAI or custom code
  • Identity and access teams who want agents in the same directory, reviews and offboarding as people
  • Risk, compliance and audit teams who have to approve agents before they touch production
  • CIOs and CISOs who need credentials and audit evidence to stay inside their own cloud

Want to try Mandate on a real system?

Request early access