Loading…
Product / CanOS · AI control plane
Not the keys.
AI can research, create, code and increasingly operate your business. CanOS gives it controlled access to the systems it needs — with permissions, approval gates and a complete record of every action.
One request · three possible answers
Every AI request resolves to exactly one of these, before anything touches your systems.
Capabilities
The question is what you’re willing to let it do.
01
Permissions
Only expose what’s needed.
02
Approvals
Keep humans where they matter.
03
Audit
Know exactly what happened.
How it works
A request passes seven checks inside one transaction. Here is one worked example, in full.
Request · approve-draft
The seven run in order inside one serialised transaction, and stop at the first failure. A deny at step 5 never leaves step 4’s write committed.
The AI can request an action. It cannot grant itself permission to perform it.
None of these is built. Until they are, autonomy stops at the approval gate.
Authority
Read
search-knowledge · find-candidates · get-candidate · list-drafts
Allowed
Available to any agent acting on behalf of a member of staff. No gate, because nothing changes.
Publish
draft-article · approve-draft · unpublish-article
Approval required
Prepared in full by the agent, held at the gate, released by the owner’s confirmation.
Move money
refunds · payouts · transfers · card details
Not in the catalogue
Not denied by a rule. Absent — there is no such action for the agent to call.
A capability you haven’t enabled doesn’t exist in the package the agent runs.
Trust boundary
What you’re usually asked to accept
What CanOS does instead
If the AI isn’t permitted to call it, there is nothing for it to bypass.
The action declares which credential it needs; CanOS resolves it server-side at execution, and it never enters the model’s context or the logs.
Model independence
In developmentYour rules shouldn’t.
CanOS
identity · authority · policy · audit
Your business
Production runs Anthropic models today. Multi-provider routing is in development.
CanOS owns the authority. AI performs the work.
Verification
Controlled rolloutA completed task is a claim, not a fact. This runs on Canarlo’s own delivery today, which is how we know it holds; it comes with the managed service rather than with the kernel.
When it goes wrong
Every vendor tells you how their system behaves when it works. This is the other half.
01
You can reconstruct it exactly
Every call, the decision at every gate, the approver who released it and what came back — appended in order, never rewritten.
02
Nothing half-finished survives
One serialised transaction per request. A failure at any check rolls the whole thing back, so there is no partial write to reconcile.
03
The blast radius was set in advance
Spend ceiling, rate limit, and a catalogue that never contained the dangerous action in the first place. The worst case is bounded before anything runs.
04
What you cannot do yet
Stop everything with one switch, or roll back a sequence that has already committed. Both are being built. Until they exist, the ceiling is the containment — which is why the catalogue stays small and money stays off it.
Pricing
Managed AI
From £3,000/month
Canarlo operates it. We design the policy, hold the integrations and run the approval queue with your team.
CanOS Platform
Talk to us
Your team operates it. In development with a small number of design partners, priced per engagement.
Already running software that needs looking after rather than AI that needs governing? Managed Operations is £500/month.
Talk to us
CanOS lets you increase that answer safely. Start with approvals everywhere, automate the actions you learn to trust, and keep humans in control of the ones that matter.