Dify

Dify's OpenAI-API-compatible model provider, in Chat completion mode.

Plugboard options

Give Dify its own key. If Dify runs in Docker, plugboard must bind a reachable address (Serving other machines):

{
  "options": {
    "listen": { "host": "0.0.0.0", "port": 4010 },
    "keys": { "dify": "{env:PLUGBOARD_KEY_DIFY}" },
    "models": ["anthropic/claude-sonnet-4-6"],
    "aliases": { "fast": "anthropic/claude-sonnet-4-6" },
    "apps": { "dify": { "allowedModels": ["fast"] } }
  }
}

Connect

Integrations → Model Provider: install OpenAI-API-compatible, then Add Model with type LLM:

Field	Value
Model Name	fast (or what you want Dify to show)
model name for API endpoint	fast, a model the dify app may use
API Key	the dify key
API Base URL	http://host.docker.internal:4010/v1 (with /v1)
Completion mode	Chat
Model context size, Upper bound for max tokens	the model's real values (the defaults are 4096)
Function Call Type	Tool Call, so agents and tool nodes can call tools
Vision Support	Support, for a model that takes images
Structured Output	Support only if you type the schema in the wrapper form (below)

Leave the other fields at their defaults. Saving sends a short test request, so plugboard must be reachable from the Dify server.

What works

Chat, streaming with usage, tool calls (Dify runs the tools), and images (Dify sends them as data: URLs).

Structured output

With Structured Output: Support, type the schema as {"name": "...", "schema": {...}, "strict": true}. A bare schema ({"type": "object", ...}) answers 400 response_format.json_schema.schema must be a JSON Schema object.

Limits

• Keep Video, Audio and Document Support off: the plugin sends them in shapes plugboard refuses (400 invalid_image). PDFs work as file parts.
• Add only LLM models. Plugboard has no embeddings, rerank or speech.
• A workflow with more than 64 tools gets 400 too_many_tools; raise maxTools.
• Dify Cloud can't reach a plugboard on your LAN; use a self-hosted Dify.