# Provider catalog

Every built-in provider slug with its upstream, request surfaces and credential header, how to add custom providers, and which CLI commands know each one.

The provider is the first path segment of the base URL: `https://guardrails.vulnes.is/{slug}/{org}/v1`. The catalog is a database table; the live list for your organisation, with your policy and key state, is printed by `vulnetix ai-firewall get`.

## Built-in providers

| Slug | Provider | Upstream base | Surfaces | Credential sent upstream |
|---|---|---|---|---|
| `openai` | OpenAI | `https://api.openai.com` | `chat`, `responses` | `Authorization: Bearer {key}` |
| `anthropic` | Anthropic | `https://api.anthropic.com` | `messages` | `x-api-key: {key}`, `anthropic-version: 2023-06-01` |
| `openrouter` | OpenRouter | `https://openrouter.ai/api` | `chat` | `Authorization: Bearer {key}` |
| `groq` | Groq | `https://api.groq.com/openai` | `chat` | `Authorization: Bearer {key}` |
| `mistral` | Mistral | `https://api.mistral.ai` | `chat` | `Authorization: Bearer {key}` |
| `deepseek` | DeepSeek | `https://api.deepseek.com` | `chat` | `Authorization: Bearer {key}` |
| `xai` | xAI | `https://api.x.ai` | `chat` | `Authorization: Bearer {key}` |
| `together` | Together AI | `https://api.together.xyz` | `chat` | `Authorization: Bearer {key}` |
| `fireworks` | Fireworks AI | `https://api.fireworks.ai/inference` | `chat` | `Authorization: Bearer {key}` |
| `alibaba` | Alibaba Cloud | `https://llm-h3tttdfc5ndvt4v4.ap-southeast-1.maas.aliyuncs.com/compatible-mode` | `chat` | `Authorization: Bearer {key}` |
| `moonshot` | Moonshot AI | `https://api.moonshot.ai` | `chat` | `Authorization: Bearer {key}` |
| `minimax` | MiniMax | `https://api.minimax.io` | `chat` | `Authorization: Bearer {key}` |
| `kiro` | Kiro | `https://q.us-east-1.amazonaws.com` | `kiro` | A Bearer access token minted per request from a stored AWS sign-in |

The upstream path is the base plus `/v1/chat/completions`, `/v1/responses`, `/v1/messages`, `/v1/messages/count_tokens` or, for Kiro, `/generateAssistantResponse`.

Every provider needs a stored key before it proxies anything (`403 provider_key_missing`). A provider Vulnesis has switched off for everyone answers `403 provider_disabled`.

> [!CAREFUL] Only openai serves the Responses API
> Codex CLI accepts no wire format but Responses, so it can only be pointed at `openai` (or a custom provider registered with the Responses protocol). Claude Code and Claude Desktop speak only Messages, so they need `anthropic`. Calling a surface a provider does not serve is `404 unsupported_api`.

## What the CLI knows about each provider

`vulnetix ai-firewall install` and `key set` use a list built into the CLI, which records the environment variables SDKs really read. It covers nine of the thirteen built-ins.

| Slug | `install` sets | Reached by environment variable |
|---|---|---|
| `openai` | `OPENAI_BASE_URL`, `OPENAI_API_BASE`, `OPENAI_API_KEY` | yes |
| `anthropic` | `ANTHROPIC_BASE_URL`, `ANTHROPIC_AUTH_TOKEN` | yes |
| `groq` | `GROQ_BASE_URL`, `GROQ_API_KEY` | yes (not verified against SDK source) |
| `mistral`, `deepseek`, `xai`, `openrouter`, `together`, `fireworks` | an informational `VULNESIS_AIFW_<SLUG>_BASE_URL` | no: set `base_url` in code, see [Snippets](/ai-firewall/cli/snippets.md) |
| `alibaba`, `moonshot`, `minimax`, `kiro`, custom providers | nothing | no |

`ANTHROPIC_AUTH_TOKEN` is used rather than `ANTHROPIC_API_KEY` because it makes the SDK send `Authorization: Bearer`. Both header forms are accepted by the gateway.

`vulnetix ai-firewall key set <slug>` refuses a slug outside the nine with `unknown provider`, and refuses `kiro` with a pointer to `kiro login`. Store keys for `alibaba`, `moonshot`, `minimax` and custom providers on the Resolve Platform's Connect tab, or with `providers[].key` in a [policy file](/ai-firewall/reference/policy-file.md), which goes straight to the API. `policy provider <slug>` and `policy model` accept any slug.

## Kiro

Kiro has no API key. `vulnetix ai-firewall kiro login` signs in with an AWS Builder ID (default) or an IAM Identity Center start URL and stores the sign-in in the [secrets vault](https://www.vulnes.is/resolve/vault) as the `kiro` key. The gateway mints a short-lived access token from it on every request.

| Item | Value |
|---|---|
| Base URL | `https://guardrails.vulnes.is/kiro/{org}/v1` |
| Inference | `POST …/generateAssistantResponse` |
| Models | `GET …/ListAvailableModels` |
| Seeded models | `auto`, `claude-opus-4.5`, `claude-sonnet-4.5`, `claude-sonnet-4`, `claude-haiku-4.5` |
| Sign-ins accepted | AWS Builder ID, IAM Identity Center. GitHub and Google Kiro sign-ins have no refresh path and cannot be used. |
| Expired sign-in | `403 vault_kiro_relogin`; run `kiro login` again |

See [Kiro](/ai-firewall/clients/kiro.md) for client setup.

## Custom providers

An organisation on Pro or above can register its own endpoint on the Connect tab of the [AI Firewall page](https://www.vulnes.is/resolve/ai-firewall). It is addressed by its slug like a built-in and gets the same authentication, provider policy, model policy and guardrails.

| Setting | Allowed values |
|---|---|
| Protocol | `openai-chat`, `openai-responses`, `anthropic-messages`. Kiro is not available. |
| Key header | `bearer` (`Authorization: Bearer {key}`), `x-api-key` (with `anthropic-version` for Messages), `cf-aig` (`cf-aig-authorization: Bearer {key}`) |
| Base URL | `https` only, with a host that resolves to a public unicast address |
| Slug | Unique within the organisation, and not any built-in slug (`409`) or reserved path word (`422`) |
| Providers per organisation | 10 |
| Declared models per provider | 100 |

| Behaviour | Detail |
|---|---|
| Visibility | Only the owning organisation. Another organisation using the slug gets `404 unknown_provider`, the same answer as a slug that does not exist. |
| Address check | Repeated on every connection, on the address the name resolves to then, so DNS rebinding cannot reach a private, loopback, link-local, CGNAT or metadata address. A refused connection is `502 upstream_blocked`. |
| Redirects | Never followed and never relayed: `502 upstream_redirect` |
| Deleting | Not plan-gated. The provider's vault keys are revoked first; if that fails nothing is deleted (`502`). |

The key is stored afterwards through the vault, never in the provider definition.

## Model names

The gateway forwards `model` unchanged. Use the provider's own id: `gpt-4o-mini` on `openai`, `openai/gpt-4o-mini` on `openrouter`, `llama-3.3-70b-versatile` on `groq`. A wrong id gets the provider's own error, relayed as is.

> [!CAREFUL] OpenRouter is a router
> Provider policy sees one provider, `openrouter`, whichever vendor serves the model behind it. Allowing it opens every vendor OpenRouter fronts. To hold a vendor or jurisdiction boundary, deny `openrouter` or put it in allowlist mode with [model policy](/ai-firewall/policy/models.md).

## Related

- [Providers](/ai-firewall/policy/providers.md)
- [Provider keys](/ai-firewall/get-started/provider-keys.md)
- [Jurisdictional control](/ai-firewall/security/jurisdictional-control.md)
