octroi.ai
Open the app

The green light, granted with proof.

A catalog of governed actions, for humans and agents.

The human-machine boundary, adjustable action by action.

A closed catalog

Agents call typed Actions by slug. There is no shell to reach.

Signed approvals

A human approves in the Channel the team already uses.

Execution stays in your network

A dispatcher runs each Job where you place it. Its credentials never leave it.

Every step is an Event

Who authorised what, where, with which rights: a query, not an investigation.

The false dilemma

Putting agents on infrastructure today means picking a bad option.

Direct access

The agent holds SSH, kubectl or cloud credentials. The only boundary is the prompt, which no auditor can bound.

Spectator chat

The agent suggests, a human copy-pastes. Safe, but nothing is recorded, reused or scaled.

Homegrown glue

Webhook, model, pipeline. Every team rebuilds the same plumbing, with no shared model of who may do what.

What is missing is the governed execution layer between agents and infrastructure.

Safety in the model, not in the prompt

The agent is a user.

Not a plugin, not a black box: a Principal under the same rules as its human colleagues. The boundary lives in the domain model and the network topology, where a prompt cannot move it.

An agent is a Principal

Its own identity, never the human's who set it up. A scope of Platforms and Environments, a budget, an autonomy level.

A closed catalog of typed Actions

Each Action is a reviewed script with a contract: inputs, risk, reversal, approval. The catalog is the whole attack surface.

Signed approval in your Channel

The approval request arrives in your messaging tool with the justification file. The decision is signed by the human identity.

How a Job moves

Propose.

A member or an agent proposes a Job from the catalog. An agent attaches its justification file: trigger, diagnosis, evidence, confidence.

Every state is named.

One lifecycle for human and agent work, and named states off the happy path too.

The path of a Job that succeeds

  1. PROPOSED
  2. APPROVED
  3. QUEUED
  4. CLAIMED
  5. RUNNING
  6. SUCCEEDED

Approve.

The effective requirement is the strictest of every policy layer: automatic, one human, two people, or deny.

Dispatch.

A dispatcher placed in your network claims the Job and runs it. It exposes no API.

Audit.

Every transition is an Event on an append-only bus, attributed to a Principal.
Autonomy is earned

Autonomy is demonstrated, not declared.

A remediation moves from free-form diagnosis to a cataloged Action, then to proposals a human approves, and only then to automatic approval. octroi.ai measures each step.

Autonomy levels, set by a human

  1. 0observe

    Proposes read-only Actions. A human approves each one.

  2. 1assist

    Read-only Actions are approved automatically. Anything that changes state waits for a human.

  3. 2operate

    Reversible Actions are approved automatically outside protected Environments.

  4. 3operate+

    Reversible Actions are approved automatically everywhere, with a change reference.

  • Destructive Actions always wait for a human, at every level.
  • Only a human raises a level, and the change is an Event.
  • A failed Job the agent had approved automatically lowers its level by one.

Questions

Short answers to what infrastructure and security teams ask first.

All questions

The basics

Is it useful without AI agents?

Yes. Members use the same catalog, the same approvals and the same audit trail. Agents come later, as Principals under the same rules, when you decide.

Security

Does the agent hold credentials to my infrastructure?

No. An agent holds a token that lets it propose Jobs within its scope and budget. The credentials that reach your systems belong to the dispatcher, which runs inside your network and exposes no API.

Where are secrets stored?

With your dispatcher, in your own infrastructure, never in octroi.ai. The dispatcher hands them to the Job at run time and masks them in logs and outputs.

What stops a manipulated agent from doing damage?

The prompt is not the boundary. An agent can only propose Actions from the catalog, within its scope, and the state machine decides what waits for a human. A manipulated agent can propose; it cannot approve its own Job or step outside the catalog.