AI Server has no plugin API that runs your code inside the server. That is deliberate: the server stays small, secure and upgradeable, and your code stays yours. You extend it from the outside, through stable interfaces.
Supported integration points
| You want to… | Use |
|---|---|
| Add AI to an application | The OpenAI-compatible API with any OpenAI SDK |
| Build agents that call your systems | Tool calling with your agent framework; your tools run in your code |
| Search your own documents | Embeddings and your vector store, or the doc-qa sample |
| Run a model that is not in the catalogue | Any GGUF model through Local AI Engine (llama.cpp) — see engines |
| Use a cloud model alongside local ones | A provider on the server's Providers page; clients use one API |
| Monitor and alert | Prometheus metrics and the usage API |
| Feed security tooling | Signed audit exports and daily shipping to your SIEM |
| Govern use centrally | Governance policy files, written by your configuration management |
| Manage servers across an organisation | AI Admin Console enrollment |
| Give staff ready-made AI apps | AI Suite apps and AI Client, which use the server automatically |
Configuration as code
Everything an operator sets in the Windows app is in files you can manage with your own tooling:
- environment variables and flags for containers and services (configuration reference);
gateway.jsonfor the worker pool;governance/*.jsonfor policy;- the settings export in Settings → Backup & restore, which you can keep in version control and import on other servers.
Changes to files apply when the server restarts; rolling restarts behind a gateway make that invisible to users.
Wrapping AI Server behind your own API
Many teams put a thin service of their own in front of AI Server — to add their own authentication, prompt templates, retrieval or logging policy — and give AI Server keys only to that service. Use one AI Server key per service so usage and limits stay attributable.
What is not an extension point
- Internal endpoints used by the Windows app and AI Suite apps, and the files in the data folder other than those listed above, can change between releases.
- Engines' own ports and APIs on the server's loopback interface are internal.
- The legacy local-AI API ends on 2026-12-31.
Partners
Integrators and software vendors who build on AI Server can join the partner programme for technical support, development licences and listing.