> ## Documentation Index
> Fetch the complete documentation index at: https://docs.mycelium-ai.co/llms.txt
> Use this file to discover all available pages before exploring further.

# Airgapped

> Fully airgapped installs for regulated and classified environments

For environments without internet egress (classified, defense, certain financial-services backbones), the runtime ships with an airgapped Dockerfile and a runbook for fully offline operation.

For the procurement framing, see the airgapped section at [myceliumai.co/byoc#airgapped](https://myceliumai.co/byoc#airgapped).

## What changes vs BYOC

* **No telemetry egress.** Crash reports, usage telemetry, and update checks are disabled at build time.
* **Bundled model fallbacks.** Local-only model adapters (Ollama, vLLM, MLC) are first-class. The runtime never assumes a hosted model is reachable.
* **Bundled connector adapters.** Connectors that require outbound calls (Slack, Salesforce, etc.) are disabled in the airgapped image; webhook receivers operate inside the customer's network only.
* **Manual update path.** New images arrive on a media drop. The customer's update cadence is decoupled from vendor release cadence.

## Bundled stack

* Runtime container (airgapped build)
* Bundled embedding model (sentence-transformers / instructor-large)
* Local LLM fallback adapter (Ollama wrapper)
* Bundled vector store (FAISS)
* Bundled audit forwarder for syslog / Splunk-compatible HEC

## Procurement notes

Common questions from customers running airgapped:

* **Where does the embedding model live?** Bundled in the image. No outbound calls.
* **How are updates delivered?** Signed media drop. The customer verifies the signature before installing.
* **What if a CVE drops between drops?** Vendor publishes the patch and an out-of-band advisory. Customer decides timing.
* **Can the customer fork?** Yes. The airgapped build is reproducible; customer can rebuild from the source repo.

## Status

The airgapped Dockerfile ships in the runtime repo. The bundled-model adapter set is stable for sentence-transformers + Ollama. Additional model adapters (vLLM, MLC) are on the roadmap.

For an airgapped install scoping conversation, contact us through the marketing site.