Resources · Models
Models
AAES doesn’t provide models. Your company connects its own accounts. Connect a model vendor, a cloud platform, a gateway or your own servers, and AAES Govern’s model gateway checks and records every call your agents send through it.
Why route model calls through Govern
In standalone Operate, an agent calls your model account directly when both use the same API, such as Claude Code with an Anthropic Messages endpoint or Codex with an OpenAI Responses endpoint. Other sources are available through AAES Govern. Through Govern, each call gets the same checks as any other action.
- Model policyPolicy decides each model call before it goes out.
- Spend limits before each callGovern reserves the estimated cost in a model budget before the call goes out.
- A sealed record of each callSettled tokens and cost go on a record you can check offline.
- Key custodyProvider keys stay in Govern. Agents hold only a Govern-issued token.
- Protocol translationGovern translates Anthropic Messages and OpenAI Responses calls for sources that accept only chat completions. A Codex agent that uses Govern’s MCP tools needs a source that accepts OpenAI Responses calls, because the translation to chat completions can’t carry the way Codex sends those tools.
- One control pointThe same checks cover every call routed through Govern, from agents inside and outside AAES Operate.
Model sources
- 8 sources tested live
- 9 sources tested against recorded responses
- 12 sources not yet verified
“Tested live” means a real call through Govern’s model gateway to the real endpoint passed, with a recorded receipt. It checks the gateway’s route, adapter, authentication, metering and spend ledger, not a full deployment. “Tested against recorded responses” means a repository test in the source’s documented format. “Not yet verified” means the preset follows the vendor’s documentation and AAES has no test yet. A source shows its best result across its endpoints and call formats, and each row lists the result for each format. Test each source with your own agents before you rely on it.
Model vendors 14
| Source | Direct in Operate | Through AAES Govern | Verification |
|---|---|---|---|
| OpenAI | Yes | Yes | Tested against recorded responses
|
| Anthropic | Yes | Yes | Tested against recorded responses
|
| xAI Grok | Yes | Yes | Not yet verified
|
| Google Gemini (AI Studio) | Yes | Yes | Tested live
|
| DeepSeek | Yes | Yes | Tested live
|
| Moonshot Kimi2 endpoints | Tested live | ||
| Global | Yes | Yes | Tested live
|
| China | Yes | Yes | Not yet verified
|
| Z.AI GLM3 endpoints | Tested live | ||
| Z.AI API | Yes | Yes | Tested live
|
| GLM Coding Plan | Yes | Yes | Not yet verified
|
| Zhipu, China (open.bigmodel.cn) | Yes | Yes | Not yet verified
|
| Qwen on Alibaba Cloud Model Studio4 endpoints | Not yet verified | ||
| Singapore | Yes | Yes | Not yet verified
|
| Beijing | Yes | Yes | Not yet verified
|
| US (Virginia) | Yes | Yes | Not yet verified
|
| Workspace host | Yes | Yes | Not yet verified
|
| Mistral AI | Yes | Yes | Tested against recorded responses
|
| Cohere (OpenAI compatibility API) | Yes | Yes | Not yet verified
|
| Groq | Yes | Yes | Tested against recorded responses
|
| Together AI | Yes | Yes | Tested against recorded responses
|
| Fireworks AI | Yes | Yes | Tested against recorded responses
|
| Meta Model API (Muse Spark) | Yes | Yes | Not yet verified
|
Cloud platforms 3
| Source | Direct in Operate | Through AAES Govern | Verification |
|---|---|---|---|
| Azure AI Foundry3 endpoints | Not yet verified | ||
| Azure OpenAI (v1 API) | No | Yes | Not yet verified
|
| Foundry Models (OpenAI v1 route) | No | Yes | Not yet verified
|
| Claude on Azure AI Foundry | Yes | Yes | Not yet verified
|
| Amazon Bedrock2 endpoints | Tested against recorded responses | ||
| bedrock-runtime | Yes | Yes | Tested against recorded responses
|
| bedrock-mantle | No | Yes | Not yet verified
|
| Google Vertex AI2 endpoints | Tested live | ||
| Regional | No | Yes | Tested live
|
| Global endpoint | No | Yes | Tested live
|
Gateways 5
| Source | Direct in Operate | Through AAES Govern | Verification |
|---|---|---|---|
| OpenRouter | Yes | Yes | Tested live
|
| Cloudflare AI Gateway7 endpoints | Tested live | ||
| To OpenAI | No | Yes | Not yet verified
|
| To Anthropic | No | Yes | Not yet verified
|
| To DeepSeek | No | Yes | Tested live
|
| To xAI | No | Yes | Not yet verified
|
| To Mistral | No | Yes | Not yet verified
|
| To OpenRouter | No | Yes | Tested live
|
| Unified chat endpoint | No | Yes | Tested live
|
| Kong AI Gateway | No | Yes | Not yet verified
|
| LiteLLM proxy | Yes | Yes | Tested against recorded responses
|
| DigitalOcean serverless inference | Yes | Yes | Tested live
|
Self-hosted servers 7
| Source | Direct in Operate | Through AAES Govern | Verification |
|---|---|---|---|
| vLLM | Yes | Yes | Not yet verified |
| SGLang | Yes | Yes | Not yet verified |
| Ollama | Yes | Yes | Not yet verified |
| LM Studio | Yes | Yes | Not yet verified |
| Hugging Face TGI | Yes | Yes | Not yet verified |
| llama.cpp llama-server | Yes | Yes | Not yet verified |
| NVIDIA NIM for LLMs | Yes | Yes | Tested against recorded responses |
“Direct in Operate” means standalone Operate can connect at least one kind of agent to the source without Govern. It is not a test result. In AAES Operate + Govern, every model call from a Claude or Codex agent goes through Govern. Hugging Face has archived TGI; its preset stays in the catalog. A logo identifies a source and claims no partnership or endorsement.
Agents that use their own models
Only standalone Operate admits these agents, after an operator turns them on. AAES hasn’t run them in its acceptance tests. They take no base URL, so their model calls go to their own vendors and Govern doesn’t govern them.
- CursorStandalone Operate only. Model calls not governed.
- Meta’s Muse agentExperimental. Standalone Operate only. Model calls not governed.
- GitHub CopilotStandalone Operate only. Model calls not governed.
Microsoft Copilot Studio is a different product. Govern registers Copilot Studio agents and governs the actions they route through it.
No model account yet?
Start with one source. OpenRouter is the fastest way in. Your own cloud platform suits enterprise procurement and data residency, and a self-hosted server suits strict data rules.
How the model gateway works · Developer details · AAES Operate