FlexAI model provider for OpenClaw: open-weight chat and vision models with live discovery, plus image generation, speech, transcription, image understanding, embeddings and video generation.
@flexai/openclaw-provider
FlexAI model provider plugin for OpenClaw. FlexAI serves open-weight models through an OpenAI-compatible API. One plugin covers chat (with image input on vision models), image generation, speech, audio transcription, image understanding, embeddings and video generation.
| Capability | FlexAI model | OpenClaw surface |
|---|---|---|
| Chat, vision chat | 24 bundled rows plus live discovery | flexai/<model> |
| Image generation | FLUX.1-schnell | openclaw infer image generate |
| Speech (TTS) | Kokoro-82M | openclaw infer tts convert |
| Audio transcription | whisper-large-v3-turbo, parakeet-tdt-0.6b-v3 | openclaw infer audio transcribe |
| Image understanding | any FlexAI vision row (default gemma-4-26B-A4B-it) | openclaw infer image describe |
| Embeddings (memory search) | bge-m3 | memory.search.provider: "flexai" |
| Video generation | MiniMax-H3 | openclaw infer video generate |
One API key covers all of them.
Install
openclaw plugins install @flexai/openclaw-provider
openclaw gateway restart
Save a FlexAI API key and pick a model:
openclaw models auth paste-api-key --provider flexai
openclaw models list --provider flexai
FLEXAI_API_KEY in the Gateway environment works too. Chat models are listed
live from FlexAI's /v1/models; the bundled catalog rows are the fallback.
Requires OpenClaw 2026.11.6 or newer. OpenClaw asks you to accept the plugin's
capabilities at install time (--accept-capabilities when scripting). See
docs/flexai.md for setup, per-capability configuration, which
models take images, discovery, the bundled catalog and per-model notes.
Vision
These bundled chat rows accept image input (input: ["text", "image"]):
DeepSeek-V4.1-Flash-1017, gemma-4-26B-A4B-it, gemma-4-31b-it,
GLM-5.3-Flash, MiMo-V2.6-Flash-RL, MiniMax-M3.1-Flash, Muse-Glimmer-30B,
Qwen3.5-9B, Qwen3.6-27B-FP8, Qwen3.6-35B-A3B-FP8, Qwen3.8-27B,
Qwen3.8-Flash-Next, plus the OCR model PaddleOCR-VL. Attach an image to an
agent turn on any of them, or run openclaw infer image describe --file <path>.
Using a different FlexAI endpoint
To point the provider at a different FlexAI endpoint, declare the provider
block under models.providers.flexai in your OpenClaw config with its baseUrl
and a models list (OpenClaw requires models for a non-bundled provider that
overrides baseUrl).
Development
npm install
npm run build
npm test
License
MIT
