Models and deployment

Bring your own model

ARIA speaks to these providers through a common interface with streaming, retries, fallback, and a circuit breaker:

• OpenAI
• Anthropic
• Google Gemini, including Vertex
• AWS Bedrock
• Any OpenAI-compatible endpoint, including self-hosted servers such as Ollama, vLLM, LM Studio, and llama.cpp

Configure a provider once, with credentials stored in encrypted configuration, and attach multiple models to it.

Not every self-hosted model supports tool calling, so validate the model you intend to use with your agents.

Managed Inference (add-on)

Managed Inference is a separately licensed add-on that provides an ARIA-supplied, self-hosted inference runtime (packaged Ollama and vLLM) with model import and model packs. Models can be imported by chunked upload, by a sneakernet inbox folder, from an internal URL, or from storage. Packs carry a manifest with per-file checksums and the license of each model.

No models are bundled: you supply the weights, which keep their own licenses.

Running without an internet connection

ARIA can run fully disconnected when you use only self-hosted models and leave external integrations, such as hosted model providers, email delivery, and security-event forwarding, switched off.

• License. ARIA licenses are signed keys verified offline against an embedded public key. The license check does not call a license server.
• Distribution. Framework and application bundles ship as checksummed archives that you can carry across a boundary, with installation scripts, and upgrade and rollback tooling.
• Dependencies. Air-gap guidance covers a local package registry or offline cache for developers.
• Prompt-injection screening. The rules scanner needs no network, and the optional screening model can be hosted inside your enclave. See Prompt-injection defense.

Plan production air-gap deployments with an ARIA engineering sizing review.

Where ARIA runs

ARIA runs on your infrastructure with Docker Compose, Kubernetes, or Node directly. It uses OpenSearch or Elasticsearch for storage and Redis for sessions, scheduling, and multi-instance operation.

The Kubernetes and Compose templates are starting points sized for development, so size production deployments to your workload.

Licensing

ARIA is licensed as a single platform license. Managed Inference and Multi-Organization are add-ons that are licensed separately. Contact ARIA for terms.

Frequently asked questions