Runware — Dify model provider plugin

Registers Runware as a first-class model provider in Dify. Runware exposes an OpenAI-compatible /v1/chat/completions endpoint, so this plugin subclasses the Dify SDK's OAICompatLargeLanguageModel and only adds Runware-specific wiring (branding, credential schema, predefined model catalog).

Install & configure

1. Install from the Dify Marketplace, or upload the .difypkg directly under Plugins → Install Plugin → Local Package.
2. Go to Settings → Model Provider → Runware and add credentials:
   • API Key — required. Get one from runware.ai.
   • API endpoint URL — optional, defaults to https://api.runware.ai/v1. Only change this if you're proxying or self-hosting an OpenAI-compatible endpoint.
3. Enable the models you want under that provider. Predefined models (Claude, GPT, Gemini, DeepSeek, Qwen, GLM, MiniMax, Kimi, Grok) show up automatically — no per-model setup needed.
4. For a model not in the predefined list, add it as a customizable model: enter the model id, the same API key, and its context size, max output tokens, vision support, and tool-calling support (check Runware's model catalog for these values per model).
5. Use the model anywhere Dify lets you pick an LLM — chatbot, agent, or workflow node.

No other connection requirements: the plugin only calls Runware's OpenAI-compatible HTTPS API (/v1/chat/completions) with the credentials above. No webhooks, no inbound connections, no other services involved.

Design note (read before changing the shape)

The Dify Plugin SDK (dify_plugin) supports exactly two configurate methods: predefined-model and customizable-model. There is no fetch-from-remote method, and there is no runtime hook that returns a credential-authenticated model list for the UI dropdown (models() / predefined_models() receive no credentials; get_ai_model_schemas is per-model). Auto-discovery from /v1/models is therefore a build step, not a runtime feature.

This plugin uses both methods:

• predefined-model — models are checked-in YAMLs under models/llm/. The API key is entered once at the provider level (provider_credential_schema).
• customizable-model — fallback for any model not in the snapshot; the user enters the model id + per-model settings (model_credential_schema).

endpoint_url and mode are required by the OAICompat base; for predefined models they are injected in RunwareLargeLanguageModel._update_credentials before delegating to the base (the DeepSeek/OpenRouter pattern).

Layout

runware/
├── manifest.yaml              # plugin manifest
├── main.py                    # entrypoint
├── pyproject.toml             # deps, dev dependency group, pytest config
├── _assets/                   # brand icons (icon.svg, icon_square.svg)
├── provider/
│   ├── runware.yaml           # methods + dual credential schemas + model glob
│   └── runware.py             # validate_provider_credentials
├── models/llm/
│   ├── llm.py                 # OAICompat subclass + credential injection
│   ├── _position.yaml         # display order
│   └── *.yaml                 # predefined models
└── tests/
    ├── test_llm.py
    └── test_provider.py

Local development

uv sync                   # installs runtime + dev deps (pytest)
uv run pytest             # run the test suite
dify plugin package ./   # produces runware.difypkg

The baseline that already works without this plugin: the official langgenius/openai_api_compatible plugin with Base URL https://api.runware.ai/v1

• API key + model id (manual, per model). This plugin adds branding, a one-time key, and the predefined catalog.

Contact, source & privacy

• Source: https://github.com/runware/dify-runware
• Privacy policy: see PRIVACY.md

Note: author in manifest.yaml must equal your Dify Marketplace organization name (your GitHub handle/org). It is set to Runware; change it if you publish under a different account.