Registries

A registry is a single upstream backend that a gateway can route to. There are two kinds:

Type	Points at	Used by
LLM	A model provider endpoint (OpenAI, Anthropic, Bedrock, …).	Chat / Responses / Messages / embeddings / images / files / rerank.
MCP	A Model Context Protocol server.	The MCP Gateway and the MCP Store.

Consumers select which registries traffic may use.

Registry screen

Open Agent Gateway → Registry.

Tab	Contents
Models	LLM providers — Available catalog cards and Connected backends. Add model / Custom Model (OpenAI-compatible custom backend).
MCP	Curated MCP servers + Custom MCP. Filter by Category.

Cards show status (Connected / Available / Active / Inactive / last test failed) and origin (Built-in / Custom). Open a card to connect, edit credentials, Test connection, or delete.

Deleting a registry

Deleting a registry is never blocked by the consumers that reference it. In the same transaction, TrustGate removes the registry from every consumer’s routing:

Consumer field	Effect
`model_policies`	The registry’s entry is dropped.
`lb_config` members	The member is removed; a pool left empty is dropped.
Smart-routing tiers	The tier is removed. Losing a middle or top tier keeps the ladder; losing the cheapest tier drops `smart_routing` and the pool falls back to round-robin.
`fallback.chain`	The registry is removed from the chain; an empty chain is dropped.
`mcp.toolkit`	Grants for that registry are removed.

The API returns 204 and logs each rewrite. Review the affected consumers afterwards: a consumer that lost its only registry, or its smart-routing floor, now behaves differently.

Connect an LLM provider

Open Registry → Models (or Getting started → Connect a provider).
Choose a provider card or Custom Model.
Enter name, credentials, and provider options (API key, Azure SP, AWS keys, base URL, custom headers, …). Optional contract pricing is on the Registry panel, not on Getting started.
Select Test connection, then Connect / Save.

You can disable a registry without deleting it when you need to take an upstream out of rotation.

Supported providers

The first-class adapters — and which of them can embed, rerank, store files, generate images, or do audio — are on Providers. Use Custom / OpenAI-compatible when the upstream speaks Chat Completions (and optionally embeddings, images, or audio) but is not a first-class card. TrustGate normalizes the inbound format (OpenAI / Anthropic / Responses / Gemini) to each provider’s wire format, so a client speaks one dialect regardless of the upstream.

Model browser

Wherever the console lets you pick models for a registry (consumer model scope, default model, smart-routing tiers), the list comes from GET /v1/models-catalog scoped to that registry:

Filter by available models keeps only the catalog models the registry’s credentials can actually list from the provider. Azure lists the resource’s deployments. If the live listing fails, the full catalog is shown. Results are cached for 10 minutes per credential.
Custom / OpenAI-compatible registries have no catalog rows, so the picker lists the models the server serves live (no price or release date).

Upstream credentials

The credential TrustGate uses to call the provider is distinct from the consumer auth your clients use:

Mode	Typical use
API key	Most providers.
Azure	Azure OpenAI (API key, service principal, or managed identity).
AWS	Bedrock (access keys or assumed role).
OAuth2	OAuth2-protected upstreams.
GCP service account	Vertex AI.

Provider credentials stay in TrustGate — applications only hold consumer keys or tokens.

Contract pricing

Set this on a connected LLM registry so Analytics Cost, telemetry cost.*, and LLM Budget dollar mode use the same rates as your contract. Open the connected card. Under Contract pricing:

Field	UI	Stored / API
List discount	Percent off catalog list (`20` = 20% off)	`discount` as a fraction (`0.2`)
Overrides	Model slug or glob (`gpt-4o`, `gpt-4o-mini*`) + input / output	USD per token

An exact slug beats a glob; among globs the most specific match wins. If a model has no override, the discount is applied to the models.dev catalog price. Otherwise the catalog price is used as-is. A model with no catalog price is unpriced: the event has no cost, and a dollar budget accrues 0 for that request. The served registry after routing is the one that applies. The same slug can have different rates on two registries. Saving empty pricing from the Registry panel clears stored rates. Getting started → Connect a provider does not set or clear them.

{
  "pricing": {
    "discount": 0.2,
    "overrides": {
      "gpt-4o": { "input": 0.0000015, "output": 0.000006 },
      "gpt-4o-mini*": { "input": 0.0000001, "output": 0.0000004 }
    }
  }
}

On update, pricing: {} clears. Omit pricing on create to leave it unset.

Connect an MCP server

Open Registry → MCP.
Pick a catalog server (one-click when no extra config) or Add custom MCP.
For custom servers, choose MCP server URL or OpenAPI document:
- MCP URL — server URL, auth None / Static header / OAuth (forwarded) with registration Automatic (DCR) or Manual (client id/secret, authorize/token URLs), plus client credentials when the server uses a token endpoint.
- OpenAPI — spec URL, optional API base URL, Validate OpenAPI, then auth None / Static header / client credentials. See OpenAPI tools.
Test the connection (or finish OpenAPI validation), then save. Open the registry to browse live tools.

See MCP for toolkits, fail mode, and agent OAuth.

Instances and the Store

A catalog MCP server on a gateway is a registry with that server’s catalog code. It can come to exist three ways:

Created by	How	Marked
An admin	Connect from the MCP tab, with the server’s credential.	Origin Built-in / Custom.
The MCP Store	Created on the first self-service install, approval, or consumer binding of a server that needs no admin credential.	`mcp_target.origin: "store"`; the console shows it as the Default instance. Deleting it is safe: the next install recreates it.
An admin, without installing	`POST /v1/gateways/{gateway_id}/registries/from-catalog` with `{ "code": "<catalog code>" }`. Idempotent; `409` when the server needs an admin credential.	`mcp_target.origin: "store"`.

Each catalog entry declares two flags, visible on GET /v1/mcp-servers-catalog next to requires_auth and url_variables:

Flag	Meaning
`self_service`	People can install it from the Store without an admin adding a credential first.
`multi_instance`	The server can be connected more than once with different configuration (a second Snowflake schema, a second API key). The registry side panel offers Add instance. For a single-instance server (one URL behind per-user OAuth) a second registry is refused with a conflict and the console explains why.

Servers whose upstream URL is a per-user template declare url_variables (name, description, required, secret, in); the Store collects those values from each user at install time.

Catalogs in the UI

The console surfaces:

Providers — supported providers, formats, and credential options.
Models — model metadata (context window, pricing, capabilities).
MCP servers — pre-seeded enterprise MCP servers you can connect in one click when no extra config is required, with their self_service and multi_instance flags.

Health checks

LLM registries can enable active health checks. Unhealthy registries are skipped by load balancing until they recover — see Load balancing.

Start here

Configure

Endpoints

Routing

Policies

MCP

Observability

Registry screen

Deleting a registry

Connect an LLM provider

Supported providers

Model browser

Upstream credentials

Contract pricing

Connect an MCP server

Instances and the Store

Catalogs in the UI

Health checks

​Registry screen

​Deleting a registry

​Connect an LLM provider

​Supported providers

​Model browser

​Upstream credentials

​Contract pricing

​Connect an MCP server

​Instances and the Store

​Catalogs in the UI

​Health checks

Registry screen

Deleting a registry

Connect an LLM provider

Supported providers

Model browser

Upstream credentials

Contract pricing

Connect an MCP server

Instances and the Store

Catalogs in the UI

Health checks