Products
AICP, the AI Control Plane
Governed access to frontier, EU-region and private AI models
AICP
AICP is one governed way for your people and applications to reach frontier, EU-region and private AI models. It runs in your environment or is hosted by us, connects to your identity provider and secrets, decides from each request's data classification where it may go, enforces budgets before money is spent, and keeps the audit evidence a compliance programme needs.
Who it is for
Organisations that already use AI across teams and applications and now have to answer where the data went, who used which model, under which policy, and what it cost.
The problem it solves
Most organisations already use AI. Few can say how, where, or at what cost.
Shadow AI
Staff paste work into AI tools nobody registered, on personal accounts. There is no inventory of who uses which model for what.
Data leaving the EU
Nothing checks whether a confidential prompt is about to reach a provider outside the EU, and nothing records it when it does.
Uncontrolled spend
Provider bills arrive after the fact. A limit that only reports overspend does not prevent it.
No audit trail
When an auditor, a DPO or an employee asks what was processed, where and under which rules, there is no record to show.
What AICP does
The model gateway underneath is replaceable. The product is identity, policy, residency routing, budgets and audit.
OpenAI-compatible API
Applications call one endpoint and address models by a stable alias, never by a raw provider model name, so the model behind an alias can change without a code change.
Routing by classification
Each request's data classification decides where it may go: a frontier provider, an EU-region provider, a private model on your hardware, or nowhere.
Classification with a ceiling
Classification is derived from the application, the user's role and the request. Content detection can only raise it, and a request above its ceiling is refused, never downgraded.
Sensitive data detection
Detectors for categories such as account numbers and leaked secrets block, redact or record, per application, category and provider. Redaction happens before the external call.
Identity and SSO
People sign in through your OIDC identity provider and software uses workload identity. Model allowlists and classification ceilings are set per team and role.
Budgets that prevent spend
Estimated cost is reserved before the upstream call and reconciled against actual tokens afterwards. Caps hold across every replica.
Usage and cost metering
Token use and cost are recorded per request and application, with dashboards and live spend in the operator console.
Audit trail and evidence
Metadata only by default, with pseudonymised users, hash-chained so an edit is detected, and exported to your SIEM. An evidence report lists models, providers, policy versions and decisions.
Operator console
An admin console behind your SSO shows the effective policy, live spend and refusals, lets operators change budgets, ceilings and detection actions, and flags drift from the reviewed baseline.
Private models on your hardware
Restricted data can be served by an open-weight model on your own GPU, so the prompt never leaves your boundary.
Residency-safe failover
When a provider fails, a request is retried only on routes policy already approved for its classification. Every attempt is recorded.
Signed supply chain
Every image is pinned by digest and signature-verified at admission. A tampered image does not start.
How it works
Every request takes the same path. What changes is where it is allowed to end.
Your people and software
- Applications and SDKs
- Employee chat interface
- Scripts and automation
AICP control plane
- Verify identity
- Derive classification
- Decide by policy
- Reserve budget
- Route and record
Where a request may go
- Frontier providersoutside the EU
- EU-region providersEU processing
- Private modelyour hardware
- Refusedabove its ceiling
- Your identity providerSSO in
- Your secret storeprovider keys
- Your SIEMaudit out
OpenAI-compatible: change the base URL
Existing OpenAI SDKs and tools keep working. Point them at AICP, authenticate with the token your identity provider issues, and call a model by its alias. Streaming responses are supported.
- No provider SDK in your codebase
- Models change behind the alias, not in your code
- A refused request returns a clear error, not a silent fallback
from openai import OpenAI
client = OpenAI(
base_url="https://aicp.example.com/v1", # your AICP endpoint
api_key=token, # issued by your identity provider
)
reply = client.chat.completions.create(
model="eu-only", # an alias, not a provider model
messages=[{"role": "user", "content": "Summarise this contract"}],
)Deployment options
In your environment
AICP runs in your own cluster, integrated with your identity provider, your secret store and your provider contracts.
- Your own provider keys
- Prompts do not pass through our infrastructure
- We deploy it and can operate it for you
Hosted by us
We run AICP for you on our infrastructure, hosted in the EU by default, in Germany, France or Bulgaria.
- Your own provider keys, if you prefer
- We act as your processor under a data processing agreement
- US hosting on request
Managed operations
For either option, we handle the work that recurs once AI is in daily use.
- Provider, pricing and model changes
- Policy tuning and budget review
- Upgrades, evidence reports and access reviews
Governance and compliance
AICP is compliance-enabling infrastructure. It enforces and documents which data reaches which provider, and gives your DPO the evidence. Whether a deployment complies remains their judgement.
- Show that no confidential prompt reached a provider outside the EU, from the record rather than from a policy document
- Verify that the audit trail was not edited before it is handed to an auditor
- Answer an employee's access request from the pseudonymised trail, without storing prompt content
- Tie every decision to the policy version that was in force when it was made
Content detection is not DLP. It can miss and over-match, which is why it may only raise a classification. The dependable boundaries are the classification ceiling, the provider allowlist and network isolation.
Planned
On the roadmap and not available yet.
- Planned
AI Act documentation from your deployment
Documentation generated from the configuration and the audit trail, with gaps marked rather than filled.
- Planned
Governed image generation
Images produced on your own GPU under the same identity, classification, budget and audit path as chat.
- Planned
Anthropic-compatible endpoint
A second API shape, for tools that speak it.
How it fits with our services
The teams behind our services build and run this product. These are the engagements it connects to.
- Private LLM Hosting and AI IntegrationPrivate LLMs on our EU infrastructure or yours, behind an OpenAI-compatible API with token-based plans, integrated into your apps, workflows and data.Learn more
- Open-Source Integration ServicesWe integrate and operate Nextcloud, Keycloak, Authentik, GitLab, Mattermost, Matrix, Odoo, Grafana and the rest of your open-source stack, with SSO and EU hosting.Learn more
Questions
Is AICP just an AI gateway?
No. A gateway routes requests to models. AICP decides whether a request may go to a given provider at all, from its data classification, the user and the application, enforces the budget before the call, and keeps the evidence. The gateway underneath is replaceable.
Does using AICP keep our data on premises?
Only on the private-model path. A request routed to an external provider still leaves your organisation; AICP restricts, governs and records that transfer. Restricted data can be confined to a model on your own hardware.
Does it make us GDPR or AI Act compliant?
No product can. AICP enforces your rules and produces the evidence a compliance programme needs. It cannot supply a legal basis or correct an unlawful use case, and the compliance judgement stays with your DPO.
Are prompts stored?
Not by default. The audit trail records metadata: who, which application, which classification, which provider, what was detected and what it cost, with users pseudonymised.
Can we use our own provider contracts and keys?
Yes. In your environment AICP uses your provider accounts, with keys held in your secret store. When we host it, you can bring your own keys too.
Which identity providers does it work with?
Any OIDC identity provider, such as Microsoft Entra ID or Okta.
AICP
Questions about a product?
Ask us about access, fit for your team or how it connects to your systems.