Dify

Dify is a platform for building LLM apps — chatbots, workflows and agents — in a visual editor. Its OpenAI-API-compatible model provider accepts a custom base URL, so your stack shows up in every model picker in the workspace.

Prerequisites#

• A key from your stack's API keys page. See Authentication.
• The id of the model you want to call — see Models. The examples below use qwen/qwen3.5-9b.
• Workspace admin rights in Dify, to add a model provider.

Add the model#

1. Open the model providers

   Click your avatar → Settings → Model Provider. In recent versions the same page is under Integrations → Model Provider.

2. Install OpenAI-API-compatible

   Since Dify 1.0 providers are plugins. If OpenAI-API-compatible isn't listed, install it from the Marketplace section on the same page.

3. Add a model

   On the provider's card, click Add Model and fill in:

   Field	Value
   Model Type	LLM
   Model Name	qwen/qwen3.5-9b
   API Key	Your Stac key
   API Base URL	https://api.trystac.com/v1
   Completion mode	Chat
   Model context size	Raise it — Dify's default of 4096 is far too small
   Function Call Type	Tool Call, if the model supports tool calling; otherwise leave Not Support

   Leave the other fields on their defaults. Older versions of the plugin label the URL field API endpoint URL.

4. Save

   Dify validates the credentials with a test call before saving. A saved model means the key and URL work.

Example#

Create a Chatbot app, pick qwen/qwen3.5-9b in the model selector at the top right of the editor, and send a message in the Debug and Preview panel:

You:        Name three primary colors.
Assistant:  Red, blue, and yellow.

The request shows up under Usage & Logs in the dashboard, the same as any other client.

Limitations#

• The model name must be on your plan. Type the exact value from Models in Model Name; a model your key's plan doesn't include returns 403.
• Agents need tool calling. Agent apps and agent nodes only call tools when Function Call Type is set and the model supports tool calling.
• Long blocking calls time out. Streaming apps are fine; a workflow node that waits for the full answer fails with 503 if generation runs past roughly 90 seconds.
• Instructions turn off automatic retrieval. When the app has instructions, Dify sends them as a system message, and a key left on Automatic then skips the knowledge base. Set it to Enabled on the key Dify uses.
• One completion per request, logit_bias is dropped, and rate limits are per key. See Differences that matter.

Next steps#

• OpenAI compatibility — what the provider is talking to, and what differs from OpenAI.
• Knowledge base — when to use Stac's retrieval instead of Dify's knowledge feature.
• Errors & rate limits — which status codes to retry.