What we plan to connect, and what has been shown.
Integration pages tend to be walls of logos. This one is a table with an evidence level, because a name on a page tells you nothing about whether it works.
Evidence ladder
Every integration must earn its label.
- E0 / Where things are now
Assessed
We have read the documentation and identified where it could fit.
“Evaluating”
- E1
Adapter exists
An adapter runs in a test.
“Experimental adapter”
- E2
Benchmarked
A reproducible benchmark is published, with its method.
“Tested with [version]”
- E3
Supported
In the product, documented and usable by customers.
“Supported”
Planned compatibility
The list.
| Technology | Where it would fit | Category | Target | Status today |
|---|---|---|---|---|
| OpenAI-compatible endpoints | One adapter for any endpoint that speaks this API shape. | Inference | Phase 1 | Planned |
| Direct model provider APIs | Native adapters for commercial model providers, using your own keys. | Models | Phase 1 | Planned |
| Open-source models | Reached through any compatible self-hosted or managed endpoint. | Models | Phase 1 | Planned |
| Enterprise-hosted inference | Models behind your own network boundary, reached through a private connector. | Inference | Phase 4 | Research |
| NVIDIA NIM | Containerized inference microservices, reached through the OpenAI-compatible adapter. Production use needs the customer's own NVIDIA AI Enterprise license. | Inference | Phase 3 | Evaluating |
| NVIDIA Nemotron | Open models as a selectable route for planning and tool-calling steps, once benchmarked. | Models | Phase 3 | Evaluating |
| NVIDIA NeMo Guardrails | An optional content and tool-call check inside a Guard decision. It complements Mandates and policy. | Safety | Phase 3 | Evaluating |
| NVIDIA NeMo Relay and Helix | Candidates for agent instrumentation and for evaluation over exported records. | Evaluation and telemetry | Phase 3 | Evaluating |
| NVIDIA NeMo Fabric | A candidate route for running external agent harnesses under the gateway. | Evaluation and telemetry | Phase 3 | Research |
| NVIDIA Dynamo-Triton | Formerly Triton Inference Server. For serving embedding and classification models, if a workload justifies it. | Inference | Phase 3–4 | Research |
| Model Context Protocol (MCP) | MCP client in Connect. A connected server grants nothing: each tool still passes Guard. | Protocols | Phase 2 | Planned |
| Agent2Agent (A2A) | Delegation to external agents under a narrowed Mandate. | Protocols | Phase 4 | Research |
| OpenTelemetry | Trace export through an adapter. The GenAI conventions are still in Development status. | Evaluation and telemetry | Phase 2 | Planned |
No integration listed here exists yet. All are planned, evaluated or under research. NVIDIA, NIM, Nemotron, NeMo and Dynamo-Triton are trademarks of NVIDIA Corporation. They are named here to describe planned technical compatibility. FLEXVOTIX has no partnership, certification or endorsement from NVIDIA.
Rule of the architecture
The core depends on none of them.
Every integration sits behind a stable interface. The release test is that the core builds, deploys and passes its full suite with any optional adapter removed, NVIDIA's included.
What AEGIS adds
Models answer. Inference serves.
Neither knows what this agent may do on whose behalf, how to resume after a crash, or how to prove what happened. That is the layer AEGIS is designed to provide, on top of whichever models you choose.
Need a particular integration first?
Requests from early access teams shape the order of work.