Request a Briefing

Hermes reference

Edge and offline deployment

Hermes is packaged to be carried into a disconnected network and installed on a single Linux host, running local models with no dependence on the internet.

What is in the bundle

A Hermes release is a single archive that contains:

  • Container images for the application, OpenSearch, OpenSearch Dashboards, Redis, and (optionally) a local inference runtime
  • Compose and Kubernetes files and an install script
  • A SHA256SUMS manifest so you can verify the bundle after transfer
  • Documentation

The install script loads the images from the archive, asks how you want to run inference, and writes the configuration. Self-extracting upgrade scripts and a rollback script are provided, with x86-64 and arm64 builds selectable.

Local models

Hermes runs with local models served by an OpenAI-compatible runtime such as Ollama or vLLM. No model weights ship in the bundle: you supply them, so you control provenance and license terms. A model-pack script builds a checksummed archive, with an optional GPG signature, for carrying models across a boundary. For the wider model layer, see models and deployment.

Offline mode

Turn on offline mode and Hermes refuses to connect to anything that is not local.

  • A process-wide guard blocks outbound connections to non-local addresses, including DNS-resolved names, and logs each refused host. Loopback, private network ranges, and hosts you list explicitly are allowed. Link-local addresses are refused.
  • Saving a hosted model provider, a security-event destination, a model import by URL, a notification channel, or an external tool connection is refused with a clear message.
  • The interface shows an offline badge, and workflow steps that try to reach an outside host fail with a message naming it.
  • An administrator turns it on with a setting and restarts the backend.

Scope: Offline mode protects the Hermes process. It does not cover child processes or other containers on the host, so isolate the network as well.

Egress test

Release packaging runs an egress test that starts the stack, drives setup, a File Ask session, a workflow, and an upgrade restart, captures network traffic, and records any attempted connection to a non-local address.

The release fails if the test fails, unless the packager deliberately skips it, in which case the bundle carries a note that the test was not run. Check the bundle you receive for that note.

Datastores

In the stock Compose file, OpenSearch, OpenSearch Dashboards, and Redis bind to the host's loopback address, so they are not reachable from the network. An administrator can change the bind address. The application's own ports remain published, so apply your own network controls.

Reference hardware

The reference sizing below comes from ARIA deployment guidance:

Hermes reference hardware sizing
ProfileGuidance
Minimum16 GB RAM, 4 or more cores, a roughly 3B-parameter quantized model, 1–2 concurrent users
Recommended32 GB or more RAM, 8 or more cores, optional NVIDIA GPU, a 7–9B-parameter quantized model
Storage100 GB or more SSD
Operating systemLinux with Docker Engine

Analysts connect from any operating system through a browser, and nothing needs to be installed on their machines.

Designed to operate without external connections

The Hermes code does not depend on content delivery networks, web fonts, analytics services, or telemetry, and its license verifies offline. Optional features can reach outside the host if an administrator turns them on: hosted model providers, the HTTP request tool, external database tools, and security-event forwarding. Offline mode removes those options. Accreditation is separate: see below.

Accreditation

Hermes supplies controls, documentation, and engineering support for your authorization process. Accreditation belongs to the deployed system and your authorizing official. See Identity and compliance alignment.

Frequently asked questions

Can Hermes run on a laptop or small appliance?

Yes, within the reference sizing above. Larger models and more concurrent users need more memory, cores, and a GPU.

How do we update a disconnected system?

Carry a self-extracting upgrade script across the boundary. A rollback script restores the previous version.

Does Hermes need a cloud account?

No. Cloud model providers are an option, not a requirement.

Back to the Hermes overview

Planning a disconnected or edge deployment?

Contact ADS