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

# Self-hosted

> Running Henry in your own environment, against your own models.

Henry Cloud is the default. For organizations that need Henry inside their own boundary, an enterprise self-hosted deployment is available: your database, your models, connected or fully air-gapped.

<Note>
  Self-hosted Henry is scoped with our team, not self-serve. [Contact sales](https://www.usehenry.ai/contact-sales) and we will walk your security team through the architecture and the deployment that fits.
</Note>

## Your own model endpoint

A self-hosted Henry can route model calls to any OpenAI-compatible endpoint: vLLM, Ollama, TGI, LiteLLM, or an internal gateway. This is configured on the deployment, not per workspace.

| Setting                       | Meaning                                                                                          |
| ----------------------------- | ------------------------------------------------------------------------------------------------ |
| `CUSTOM_MODEL_BASE_URL`       | OpenAI-compatible base URL ending in `/v1`. Required.                                            |
| `CUSTOM_MODEL_ID`             | The model name your endpoint serves. Required.                                                   |
| `CUSTOM_MODEL_API_KEY`        | Bearer key, if your endpoint wants one.                                                          |
| `CUSTOM_MODEL_LABEL`          | Display name in Henry. Default "Self-hosted model".                                              |
| `CUSTOM_MODEL_CONTEXT_WINDOW` | Context size in tokens. Unset disables prompt-size checks.                                       |
| `CUSTOM_MODEL_EXCLUSIVE=1`    | Route every model call, including utility calls, to this endpoint. The "no data leaves" posture. |

Behavior:

* The endpoint appears in **Settings → Models** as a single catalog entry.
* Calls go directly to your endpoint. Henry's gateway never sits in front, and there is no failover.
* No reasoning or thinking configuration is sent.
* Self-hosted inference records no Henry-side model cost.
* Use a model id that does not collide with a catalog model. If it does, calls still reach your endpoint but are priced at the catalog model's rates; `CUSTOM_MODEL_EXCLUSIVE=1` avoids this.
* With exclusive mode on, every workspace's model preference resolves to your endpoint.

## Identity

Self-hosted deployments support SAML single sign-on the same way Henry Cloud Enterprise does. See [Single sign-on](/user-guide/admin/single-sign-on).

## What to expect from the engagement

* Architecture walkthrough for your security team, including the [security architecture](/developer-guide/security-architecture) and subprocessor list.
* Sizing and deployment scoping for your environment.
* Versioned updates, including offline updates for air-gapped installs.