octroi.ai
Open the app

All posts

Introducing octroi.ai

Teams that want AI agents on their infrastructure face a false choice. Give the agent direct access, and the only thing between it and production is a prompt. Keep it in a chat window, and a human copy-pastes every suggestion with no trace left behind. Glue a webhook to a model and a pipeline, and every team rebuilds the same plumbing with its own rules.

octroi.ai is the layer that was missing: a catalog of governed actions, for humans and agents. The human-machine boundary, adjustable action by action.

The agent is a user

In octroi.ai, an agent is a Principal, like any member of your team. It has its own identity, never that of the person who set it up. It has a scope of Platforms and Environments, a budget, and a closed catalog of Actions it may propose.

An Action is a reviewed, typed script identified by a slug such as service.restart. Its contract says what it means: inputs, risk, reversal, approval. The catalog is the whole attack surface, and your team owns it.

Every Job has a path

A Job moves through one lifecycle, whoever proposed it:

PROPOSED → APPROVED → QUEUED → CLAIMED → RUNNING → SUCCEEDED

An agent’s proposal carries its justification file: the trigger, the diagnosis, the evidence and its confidence. The effective requirement — automatic, one human, two people, or deny — is the strictest of every policy layer, and no layer can lower it. Humans approve in the messaging Channel they already use, and the decision is signed by their identity.

A dispatcher placed in your network then claims the Job and runs it. It exposes no API, and the credentials it uses never leave it.

Autonomy is earned

An agent starts by observing. As its proposals are approved and its Jobs succeed, a human can raise its autonomy level. A failure lowers it automatically. Destructive Actions always wait for a human.

What octroi.ai promises

octroi.ai does not claim to know what your scripts do. It tells you, for every Job, who authorised what, where, and with which rights — and it is useful from the first day, with or without AI.

Open the app to catalog your first Action.