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

# Prompt & model controls

> Shape requests before they reach the model: inject and version system prompts with prompt_template, and restrict which models a consumer may call with model_allowlist.

Two policies shape the request itself before it is forwarded. Both run at `pre_request` and
are scoped like any other [policy](/trustgate/policies/overview).

## `model_allowlist`

Restrict which models a consumer or gateway may call, matching by name with glob patterns.
This is the policy-based complement to a consumer's
[`model_policies`](/trustgate/routing/model-resolution) — use it for a gateway-wide default.

| Setting                  | Type      | Default  | Notes                                                                 |
| ------------------------ | --------- | -------- | --------------------------------------------------------------------- |
| `allowed_models`         | string\[] | —        | Name patterns permitted, `*` wildcard (e.g. `gpt-5*`). **Required.**  |
| `behavior_on_disallowed` | enum      | `reject` | `reject` (403) or `substitute`.                                       |
| `substitute_with`        | string    | —        | Model written back when `substitute` (must match the allowlist).      |
| `default_model`          | string    | —        | Model injected when the request omits one (must match the allowlist). |

```json theme={null}
{
  "slug": "model_allowlist",
  "settings": {
    "allowed_models": ["gpt-4o", "gpt-4o-mini"],
    "behavior_on_disallowed": "substitute",
    "substitute_with": "gpt-4o-mini"
  }
}
```

## `prompt_template`

Inject and version prompts server-side, so system instructions live in the gateway rather
than in every client. It supports two modes, which can be combined:

* **Mode A — inject templates:** render operator-defined system prompts from context
  variables (headers or JWT claims) and inject them into the request.
* **Mode B — named templates:** maintain named, versioned templates that clients reference by
  `{template://name@label}`, with required-variable validation.

| Setting                       | Type   | Default    | Notes                                                                                                |
| ----------------------------- | ------ | ---------- | ---------------------------------------------------------------------------------------------------- |
| `template_engine`             | enum   | `mustache` | Rendering engine (`mustache` in v1).                                                                 |
| `context_variables`           | map    | —          | Named values resolved from `header` or `jwt_claim` sources.                                          |
| `inject_templates`            | array  | —          | Mode A: `[{ id, content, position, on_existing_system }]`.                                           |
| `named_templates`             | array  | —          | Mode B: named templates with `versions[]` (each `{ version, labels, content, required_variables }`). |
| `allow_untemplated_requests`  | bool   | `true`     | When false + named templates set, reject requests with no template reference.                        |
| `on_missing_context_variable` | enum   | `error`    | `error`, `empty_string`, or `skip_injection`.                                                        |
| `default_label`               | string | —          | Label used when a reference omits `@label`.                                                          |

```json theme={null}
{
  "slug": "prompt_template",
  "settings": {
    "context_variables": {
      "tenant": { "source": "header", "name": "X-Tenant-Id" }
    },
    "inject_templates": [
      { "id": "base", "content": "You are the assistant for {{tenant}}.", "on_existing_system": "merge" }
    ]
  }
}
```

Injected content is placed in the `system` role; `on_existing_system` chooses whether to
`merge` with or `replace` an existing system prompt.