OpenAI-compatible nodes

Add custom chat, Responses, Anthropic, or embedding endpoints and verify their model IDs.

A provider node is an endpoint DurinDoor does not ship as a registry provider. Use it for an internal gateway, a self-hosted vLLM, a vendor that only speaks OpenAI or Anthropic routes, or an embeddings-only server.

Open Dashboard, then Providers. The Custom Providers section has Add OpenAI Compatible and Add Anthropic Compatible. Give the node a stable prefix so clients can request prefix/model.

OpenAI-compatible

Click Add OpenAI Compatible. Name and prefix are required. Prefix is the client-facing provider token in prefix/model strings, for example oc-prod/llama-3.1-70b.

Pick Chat Completions or Responses API. Default base URL is https://api.openai.com/v1. Point it at your gateway, including the /v1 suffix the upstream expects.

Save, then open the node and add an API key connection. The connection form requires a key even if the upstream ignores it. For a local server that does not check auth, paste a placeholder such as local-dev-key.

Add model names, or enable passthrough if the node should forward unknown ids. An empty saved model list exposes no models for that node. Use Import on the node page to discover /v1/models from the upstream, then save the rows you want.

GET /v1/models on DurinDoor is the list clients should trust. Discovery against the node needs both providerSpecificData.baseUrl and a token.

Anthropic-compatible

Same Providers page, Add Anthropic Compatible. Default base URL is https://api.anthropic.com/v1. If you paste a URL that already ends in /messages, DurinDoor strips that suffix so runtime does not double it.

DurinDoor translates OpenAI-shaped client requests to Claude Messages when the node needs it. Tools, images, reasoning, and streaming do not always map. Test those before you put the node in a combo.

Third-party Anthropic gateways that are not api.anthropic.com get both x-api-key and Authorization: Bearer so either header style works.

Custom embeddings

Embedding-only nodes live under Dashboard, Media providers, embedding. Create an embedding-only node there. A trailing /embeddings on the base URL is stripped. Default base URL is https://api.openai.com/v1. Clients call /v1/embeddings.

Quota display for compatible connections

OpenAI-compatible and Anthropic-compatible connections have no fixed billing endpoint, so DurinDoor cannot show Provider Limits for them out of the box. Set providerSpecificData.quotaEndpoint on the connection (via PUT /api/providers/{id}) to describe where the quota lives and how to read it:

{
  "providerSpecificData": {
  "quotaEndpoint": {
    "url": "https://api.example.com/v1/credits",
    "method": "GET",
    "auth": "bearer",
    "headers": { "X-Org": "acme" },
    "quotas": {
      "credits": {
        "used": "$.data.used_usd",
        "total": "$.data.limit_usd",
        "resetAt": "$.data.renews_at",
        "currency": "usd"
      }
    }
  }
  }
}

auth is bearer (default), x-api-key, or none. The connection's saved API key is sent as the credential; none sends no auth header. used/total/resetAt are dot/bracket paths into the JSON response ($.a.b[0].c), not full JSONPath. A quota only shows up once total resolves; a path that cannot resolve is left out rather than reported as 0/0.

Aliases and naming

Keep the prefix lowercase and stable. Leave the upstream model id after the slash alone. Avoid spaces.

oc-lab/llama-3.1-70b
ac-internal/claude-sonnet
emb-search/text-embedding-model
coding-default

coding-default is a combo or a model alias, not a node prefix. Model aliases map a short client name to an existing model ID. They do not create an additional top-level catalog row for that short name. Combos are a fallback chain; see Combos.

Dedicated local registry ids (ollama-local, lm-studio, vllm) are usually a better fit than a generic node when one exists. Those ids omit Authorization when no key is saved. Compatible nodes still require the key field. See Local router providers and Free and local.

If DurinDoor runs in Docker, localhost inside the container is the container. Use a compose service name or the host gateway for a server that runs on the machine.

Validate the URL before saving the node. Only http and https are accepted. The probe is SSRF-guarded and stops after 10 seconds. Connection refused, DNS failure, and certificate errors come back as that 400/403 payload rather than hanging the dashboard.

Icon URLs are optional and validated before saving. A missing name or prefix is rejected with 400.

Custom nodes may reject content-part arrays, tool calls, or unknown OpenAI fields. Prefer one node per upstream behavior rather than mixing incompatible models behind the same prefix.

Before you put a node in a combo: confirm GET /v1/models lists the names you expect, send a small chat request, then test streaming and tools if your clients need them. Test images, audio, or embeddings only when those routes are in play. Usage logs show provider name, model name, latency, and errors.