All servicesTrustworthy AI agents
Your AI agents do the work. Gates check what they do, and a journal shows it.
Cerbère installs and tunes the open source Laeka integrity framework beside your AI agents. For Quebec small businesses and non-profits that hand real work to an agent and want to stay in charge.
On quote
Rules that run
A rule in a prompt is a request. A gate is a program that runs outside the model and returns yes or no.
Everything journaled
Every gate writes one line on every run, whatever the outcome, so you can see what actually ran.
Open code
The framework is released under Apache-2.0. Read it yourself, or have anyone you trust read it.
Nothing arms itself
A gate runs only because your settings file names it. That file stays yours.
Your agent says 'done'. How do you know?
An agent that writes code or runs commands can report a success that is not real, delete what it was told to keep, or install a package that does not exist. You wrote it instructions, and for a while they hold. Then the conversation gets long and the rule quietly stops applying, with no error on screen.
| Rules in the prompt | Building your own gates | With Cerbère | |
|---|---|---|---|
| What holds the agent back | A written request the model can stop following | Home-made scripts, written and tested one by one | Deterministic gates already published with their tests, tuned to your use |
| Proof the rule ran | None | Whatever you remembered to log | One journal line per run, even when nothing is blocked |
| Destructive command | It runs if the model sends it | Blocked if you thought of that case | Refused by a gate that says why and offers the working alternative |
| A broken gate | Not applicable | Often invisible: a dead gate makes no noise | A daily check flags gates that are wired but silent |
| Who sees what is going on | The agent, unless someone rereads everything | You, if you have time to reread everything | You: a read-only local panel shows gates, sessions and what is waiting on a person |
| Documentation | In the head of whoever wrote the prompt | One more thing to write | Handed over in writing: each gate, its setting, how to remove it |
The framework does not make the model smarter. It makes the model's mistakes cheap, visible and reversible.
6 cyber security agencies, including the Canadian Centre for Cyber Security, co-signed guidance on the careful adoption of agentic AI. It states: 'Privileges assigned to agents directly determine the level of risk they can introduce.'
What changes
Four things you did not have before
Gates on commands and writes
The public repository ships fifteen deterministic gates on what the agent runs and writes. We wire the ones that match your risks, not all of them on principle.
A journal you can rely on
Every pass leaves a trace. Bypassing a gate is allowed, but the journal records it: nothing gets switched off quietly.
Health checked daily
A sentinel rereads the journal and flags any gate that is wired but has gone quiet, because a dead gate looks exactly like a good day.
New rules on trial
A proposed rule faces two judges, then seven days of observation where it logs what it would have blocked and blocks nothing. Then you decide.
Who it is for
A small business coding with an agent
Your team uses Claude Code or Pi to write, test and deploy. You want a deletion or a false 'done' stopped before it reaches your repositories.
A non-profit or practice with sensitive data
You are thinking of handing tasks to an agent, but personal information is involved. You want it in writing what the agent can touch, and a record of what it did.
This is for you if
- An AI agent writes code, runs commands or changes files in your organization.
- You have already seen an agent report finished work that was not finished.
- You want a written record of what the agent did, readable by someone other than the agent.
- You would rather run open code than a black box.
Probably not for you if
- You only use a chat assistant for drafting, and it never acts on your systems.
- You want a promise that an agent will never be wrong. Nobody can keep that promise.
- You want an AI already in service tested, not fenced in: see the AI red team service.
What is included
What is included
Written scope
Which agents, which machines, which repositories, which data. Set in writing before anything is installed.
Gate selection
We keep the gates that answer your real risks and tell you which ones would be useless for you.
Install and wiring
The framework goes in beside your agent, in the version that fits: Claude Code when a person works in the session, Pi for an unattended agent.
Journal and panel
The gate journal and the read-only local panel are put into service, so you can see which gate bites, on what, and how often.
Daily check
The sentinel is set to flag any gate that is wired but silent.
Team training
How to read the journal, how to bypass a gate in the open, how to propose a new one.
Documentation handed over
Every gate we wired, its setting, why it is there and how to remove it, in writing.
How it goes
How it goes
-
Step 01
You write
Describe your agents and what they do. Written reply within 24 business hours.
-
Step 02
We scope
Scope, risks, chosen gates and quote, in writing. Nothing is installed before you agree.
-
Step 03
We install
The gates go in and are tried on a real case: one command blocked on purpose, with its line in the journal.
-
Step 04
We hand over
Documentation, training, and the settings in your hands. You can carry on without us.
Pricing
Price
-
Trustworthy AI agents
On quote
depending on the number of agents and the scope.
- Written scope and quote before any work
- Install, tuning and trial of the chosen gates
- Training and documentation handed over
- The framework is open source: no licence fee for it
The quote covers Cerbère's work. Subscriptions to your models and AI tools stay in your name, at the vendor's price, with no hidden margin.
What we do. What we don't.
We do
- Put everything in writing: scope, quote, settings, documentation.
- Leave your accounts, repositories and settings files in your hands.
- Tell you when a gate would be useless for you.
- Show you how to remove what we installed.
We don't
- Promise that an agent will never make a mistake again.
- Tie you into a contract you cannot leave.
- Wire a gate without telling you what it blocks.
- Present the framework as a certification or as compliance.
Frequently asked questions
What is the Laeka integrity framework?
A set of small mechanisms that sit between the agent and the machine: deterministic gates on what it runs and writes, rules applied at the moment of risk, and a journal showing that every gate ran. It is open source, released under Apache-2.0.
Does it make my agent safe?
No, and nobody can honestly say so. The framework does not make the model smarter. It makes the model's mistakes cheap, visible and reversible. We reduce a risk, we do not remove it.
Which agents does it work with?
The public repository provides two versions: one for Claude Code, where a person works in the session, and a stricter one for Pi, an unattended agent. If you run something else, we tell you in writing what applies and what does not, before any quote.
Will the gates get in the way of the work?
In the Claude Code version, a gate that crashes lets the work through and says so. That is deliberate: a safety net that halts everything is the one people switch off. In the Pi version the gate is the last wall, so it fails closed.
Can a gate be bypassed?
Yes. Bypassing is allowed, and the journal records it. What is not accepted is an invisible bypass.
The code is open. Why not install it ourselves?
You can: the repository and its documentation are public. Cerbère brings the choice of gates for your risks, the tuning, a trial on a real case, training, and documentation in French or English.
Are there dependencies to install?
No. The framework uses the Python standard library only: no package to add, no dependency to audit.
The framework has sixteen layers. Do you install all of them?
No. The public repository publishes the first part: the gates, the shield of behavioural rules, the memory, the rule governor and two panels. The specification of all sixteen layers is on laeka.ai. Your quote names exactly what gets installed.
Does this make us compliant with Quebec's Law 25?
No. The framework is neither a certification nor legal advice. It gives you a written record of what an agent did, which helps when someone asks. Your legal obligations remain yours to confirm with the person in charge of personal information in your organization.
What does it cost?
On quote, depending on the number of agents and the scope. You get the quote in writing before any work starts.
Tell us what your agents do.
A few lines are enough: which agents, on which machines, for which tasks. Written reply within 24 business hours, saying what applies to you and what would be a waste.
Written reply within 24 business hours. No video calls, no booking calendar.