What is Raucle?

EPIC28 Ltd · 1 September 2026

You have probably heard that AI agents are coming to the workplace. An agent is software with a brain: instead of just following fixed steps, it reads a request, decides what to do, and uses tools to do it. The tools are the ordinary systems a business already runs. A bank account. A database. A payment button. A document store.

That is convenient, and it is also the problem. When software gets to decide which tools to use, two questions follow it everywhere. Who said it was allowed to do that? And can you prove what it actually did?

Raucle is an open-source system built by EPIC28 to answer both questions. It sits between an AI agent and the tools the agent wants to use. Every time the agent tries to do something, Raucle checks the action against a set of permissions, writes down what happened, and signs the record so nobody can quietly change it later.

The rest of this post explains how it works in plain language, with three short animations.

1. The checkpoint

Think of a building site. Anyone can walk up to the gate, but only people with the right badge get in, and a guard writes down every name whether they pass or not. Raucle is that gate and that logbook for AI agents.

Before an agent can use a tool, the team that runs it defines what that agent is allowed to do: which tools, on which records, for how long. Those permissions are written into a small encrypted file called a capability, and the agent has to carry it, like a badge.

The checks are not guesses and they are not vibes from an AI model. Raucle compares the request against the capability mathematically: either the action is covered by the permission, or it is not. A call to look up a customer's record in the allowed system passes. The same agent trying to move money out of the account it was never granted? Stopped, before the tool is ever touched.

2. The record

Pass or stop, every attempt gets written down. The record says: this agent, holding this permission, tried this action, at this time, and the answer was allow or deny. Then the record is signed with the same family of cryptography that protects banking transfers, and linked to the record before it.

Why the linking matters: you cannot remove or edit one page of the logbook without breaking the seal on every page after it. Tampering becomes visible. And because each record is signed by your organisation's own key, anyone you show it to can confirm it came from you, without having to take your word for it.

3. The audit

This is where the logbook earns its keep. Months later, someone asks the question every regulated business eventually faces: what exactly did that AI system do last quarter, and was it allowed to do it?

Today, most organisations answer with ordinary log files. The trouble with log files is that they are claims. Whoever runs the system can edit them, and whoever reads them has to trust the person who handed them over. A regulator, an auditor, a court, an opposing lawyer: they all know this.

With Raucle, the answer is a set of signed, linked records that the auditor can verify with free tools, offline, without calling your vendor or your IT team. The record either checks out mathematically or it does not. That is the difference between a claim and evidence.

What it means in practice

For a bank: you can show exactly which payments an agent was authorised to make, and prove it made no others. For a hospital: the records show which patient system an agent touched and which it did not. For a manufacturer: every production change an agent made carries a signed record for the customer's audit. For a retailer: which customer records an agent read, and which it never reached. For a government department: when the National Audit Office comes asking, the evidence pack is already assembled.

Raucle is open source under the Apache-2.0 licence. You can read every line of it, run it yourself, and pay nobody for the privilege. It plugs into the major agent frameworks, and it is built to be boring: small, fast, and fail-closed, so when something goes wrong the agent stops rather than carries on.

If your organisation is deploying AI agents and the "who authorised this?" question keeps coming up in meetings, that question is what Raucle was built for. Get in touch at [email protected], or read the technical companion piece below.