Integrations
AI platform integrations
Connect models from Anthropic, OpenAI, Google and the other major providers alongside the internal ones you host yourself, and govern them through one layer.
The question that matters
Where does it run?
Not which provider. Hosting determines whether request content leaves your perimeter, and it is the first thing a compliance review asks about.
Providers
6
Inside perimeter
4
Named, not supported
A model reached over an API is governed the same way whoever serves it.
So there is no per-provider work and no support matrix. What varies is not the provider but the hosting, the pinned version and what each model was approved for — and those are properties of a model in your estate rather than of a company.
No support matrix, deliberately
A model reached over an API is governed the same way whoever serves it. The routing rule names a requirement; the provider is what satisfies it.
- Anthropic
Anthropic
Pinned to a dated snapshot, with an approved use per model rather than blanket approval.
- OpenAI
OpenAI
Governed identically whether reached through the vendor API or a private endpoint.
- Google
Google
Hosting is recorded per model, because that is the field a data-residency review asks about first.
- Meta
Meta
Open weights, typically self-hosted here. Nothing leaves the perimeter.
- Mistral
Mistral
Self-hosted in the sample estate, scoped to one approved use.
- Your own weights
Your own weights
A model you host is a governed dependency like any other: an owner, a pinned version, an approved use.
What is actually governed
A model is a dependency with an owner, a version and an approved use.
Not a setting. Every model in the estate has a human who signed off on it, a dated snapshot it is frozen on, a place it runs and a stated purpose — because approval is never blanket.
Models in the sample estate
7 models · 6 providers · 4 inside the perimeter| Model | Pinned to | Where it runs | Approved for |
|---|---|---|---|
| Claude SonnetAnthropic | 2026-04-11 | private endpoint | Customer-facing reasoning and refund adjudication |
| Claude HaikuAnthropic | 2026-04-11 | private endpoint | High-volume ticket triage and classification |
| GPTOpenAI | 2026-03-28 | vendor API | Invoice extraction and ledger reconciliation drafting |
| Gemini ProGoogle | 2026-02-19 | vendor API | CRM record hygiene and duplicate detection |
| LlamaMeta | 2026-01-30 | self-hosted | Contract reading. Nothing leaves the perimeter. |
| MistralMistral | 2026-03-05 | self-hosted | Vendor document classification, procurement only |
| Public web UIConsumer chat assistants | — | vendor API | Not approved for any action on a system of record. |
Models
7
under governance
Providers
6
distinct
Never leave the perimeter
4
self-hosted or private endpoint
Needing a decision
2
retiring or blocked
IllustrativeIllustrative figures for the OpsAI sample estate. Versions are dated snapshots this estate has frozen on, not claims about a provider’s catalogue.
Why the version is a date and not a version string
Inventing a precise vendor version would be a factual claim about somebody else’s product catalogue. A dated snapshot is a fact about this estate — the thing it froze on — and it is also how regulated organizations actually pin a model. When an incident asks which model produced a decision, the pin is the answer.
Which model answers
A governed decision, not a constant in somebody's repository.
Every agent names a model, and in most estates that name is typed into code. Making it a rule means a provider change, a residency requirement or a retirement is an edit rather than a pull request per agent.
Routing rules
6 rules| Workload | Decided by | Routes to |
|---|---|---|
| Customer-facing refund adjudication | Policy reasoningLatency budgetResidency | Claude Sonnet |
| High-volume ticket triage and classification | VolumeLatency budgetResidency | Claude Haiku |
| Invoice extraction and ledger reconciliation drafting | Structured extractionPolicy reasoning | GPT |
| Contract reading and clause extraction | ResidencyLong context | Llama |
| Vendor document classification, procurement only | ResidencyVolume | Mistral |
| CRM record hygiene and duplicate detection | VolumeStructured extraction | Gemini Pro |
Every rule names a requirement. None names a winner.
“Reasoning over a written policy, inside the perimeter, with a person waiting” is a requirement you can state and be held to. It also stays correct when the model catalogue changes, which a comparison does not.
Models needing a decision
- Gemini Proretiring
- Superseded for this use in March. Two agents still call it; both move next sprint.
- Public web UIblocked
- Reached through a browser rather than an API, so it is governed at the data boundary instead. Pasting customer data into one is the exposure this row exists to name.
Where to start
Find out which models your estate can reach, then which ones nobody approved.
The second list is usually the interesting one, and it is rarely the same as the list somebody wrote down. Discovery reads the egress proxy to find the difference.