> ## 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.

# Registries

> A registry is an upstream backend — an LLM provider endpoint or an MCP server — with its own credentials, options, and health checks.

A **registry** is a single upstream backend that a gateway can route to. It is the one
abstraction for everything downstream: there are two kinds.

| Type      | Points at                                                  | Used by                                   |
| --------- | ---------------------------------------------------------- | ----------------------------------------- |
| **`LLM`** | A model provider endpoint (OpenAI, Anthropic, Bedrock, …). | Chat/Responses/Messages traffic.          |
| **`MCP`** | A Model Context Protocol server.                           | The [MCP plane](/trustgate/mcp/overview). |

A registry has a `name`, an `enabled` flag, a `description`, and either an `LLMTarget` or
an `MCPTarget`. [Consumers](/trustgate/concepts/consumers) and
[roles](/trustgate/concepts/roles) reference registries to decide where traffic goes.

## LLM registries

An `LLMTarget` selects a provider and carries its options and upstream credential:

| Field              | Meaning                                                                                                                     |
| ------------------ | --------------------------------------------------------------------------------------------------------------------------- |
| `provider`         | One of the [supported providers](#supported-providers).                                                                     |
| `provider_options` | Provider-specific (e.g. OpenAI `{api, base_url}`; Vertex `{project, location, version}`; `openai_compatible` `{base_url}`). |
| `auth`             | The upstream credential (see [Target auth](#target-auth)).                                                                  |
| `health_checks`    | Optional active health checking for load-balancer eligibility.                                                              |

### Supported providers

`openai` · `openai_compatible` · `anthropic` · `azure` · `bedrock` · `google` (Gemini) ·
`vertex` · `groq` · `mistral` · `deepseek`.

Use `openai_compatible` (with a `base_url`) for any provider that speaks the OpenAI Chat
Completions wire format but isn't listed above.

TrustGate normalizes the inbound format (OpenAI / Anthropic / Responses) to each provider's
wire format, so a client speaks one dialect regardless of the upstream.

### Target auth

The credential TrustGate uses to call the **provider** (distinct from the
[consumer auth](/trustgate/concepts/auth) your clients use):

| `type`                | For                        | Key fields                                                          |
| --------------------- | -------------------------- | ------------------------------------------------------------------- |
| `api_key`             | Most providers             | `api_key`, optional header/param placement.                         |
| `azure`               | Azure OpenAI               | endpoint, version; API key, service principal, or managed identity. |
| `aws`                 | Bedrock                    | access key/secret + region, or an assumed `role`.                   |
| `oauth2`              | OAuth2-protected upstreams | full client-credentials config.                                     |
| `gcp_service_account` | Vertex AI                  | encrypted service-account JSON.                                     |

## MCP registries

An `MCPTarget` points at an MCP server:

| Field       | Meaning                                                                      |
| ----------- | ---------------------------------------------------------------------------- |
| `code`      | Catalog code (e.g. `com.asana/mcp`).                                         |
| `url`       | `https://` endpoint.                                                         |
| `transport` | `streamable-http` (the supported transport).                                 |
| `headers`   | Static headers to send upstream.                                             |
| `auth`      | MCP auth mode: `none` · `static` · `passthrough` · `exchange` · `forwarded`. |

The `exchange` mode supports token-exchange patterns (`impersonation`, `delegation`,
`obo`, `token_exchange`) for downstream identity. See [MCP](/trustgate/mcp/overview).

## Catalogs

TrustGate ships read-only catalogs to help you configure registries:

* **Providers** (`GET /v1/providers-catalog`) — supported providers, wire formats, auth
  types, and the options schema.
* **Models** (`GET /v1/models-catalog`) — model metadata (context window, pricing,
  capabilities), synced from OpenRouter.
* **MCP servers** (`GET /v1/mcp-servers-catalog`) — pre-seeded enterprise MCP servers.

## Managing registries

CRUD lives under `/v1/gateways/{gateway_id}/registries`. You can **test connectivity**
before saving (`POST …/registries/test-connection`) and **list an MCP registry's tools**
(`GET …/registries/{id}/tools`, which calls the live MCP server). See the
[Registries API](/trustgate/api/overview).

<Note>
  For MCP registries, the data plane must reach the upstream `url`. Private VPC MCPs need [Hybrid](/neuraltrust/deployment/deployment-models).
</Note>