| name | media-config |
| description | Configure audio transcription and image/video understanding for channels |
Media Configuration
Configure how DevClaw processes images, videos, and audio received from messaging channels (WhatsApp, Telegram, Discord, WebUI).
Settings live in config.yaml under the media: section and can also be changed via WebUI → Configuração.
Vision (Image/Video Understanding)
Controls how images and video frames are described before being added to conversation context.
media:
vision_enabled: true
vision_model: ""
vision_detail: "auto"
max_image_size: 20971520
Available Vision Models
| Provider | Model | Notes |
|---|
| Z.AI | glm-4.6v | Flagship, 128K, native tool use |
| Z.AI | glm-4.6v-flashx | Lightweight, affordable |
| Z.AI | glm-4.6v-flash | Free tier |
| Z.AI | glm-4.5v | 106B MOE, thinking mode |
| OpenAI | gpt-4o | Best quality |
| OpenAI | gpt-4o-mini | Fast, cheap |
| OpenAI | gpt-4.1 | Latest |
| OpenAI | gpt-4.1-mini | Balanced |
| Anthropic | claude-sonnet-4-20250514 | Claude Sonnet 4 |
| Anthropic | claude-opus-4-20250514 | Claude Opus 4 |
| Anthropic | claude-haiku-3-5-20241022 | Fast, cheap |
| Google | gemini-3-pro | Multimodal flagship |
| Google | gemini-3-flash | Fast multimodal |
| Google | gemini-2.5-pro | Strong reasoning |
If vision_model is empty, the main model from config is used. Set a dedicated vision model when:
- The main model doesn't support images (e.g. text-only)
- You want a cheaper model for image understanding
- You want the best vision quality regardless of chat model
Vision Detail
- auto: Let the API decide based on image size
- low: Faster, fewer tokens (~85 tokens/image)
- high: Detailed analysis, more tokens (~1590 tokens/1092×1092 image)
Audio Transcription
Controls how voice messages and audio files are converted to text.
media:
transcription_enabled: true
transcription_model: "whisper-1"
transcription_base_url: "https://api.openai.com/v1"
transcription_api_key: ""
max_audio_size: 26214400
Available Transcription Models
Choosing a Transcription Provider
- Z.AI GLM-ASR-2512: Best if already using Z.AI as main provider. Low CER, supports Chinese/English/dialects.
- OpenAI GPT-4o Transcribe: Best quality, supports diarization variant.
- Groq Whisper: Fastest (189–216x realtime), cheapest, OpenAI-compatible endpoint.
- OpenAI Whisper-1: Reliable fallback, broadest format support (SRT, VTT, verbose JSON).
Using a Different Transcription Provider
When the main LLM provider doesn't support audio transcription (e.g. Anthropic, xAI), set transcription_base_url and transcription_api_key:
api:
provider: anthropic
api_key: ${DEVCLAW_API_KEY}
media:
transcription_enabled: true
transcription_model: whisper-large-v3
transcription_base_url: https://api.groq.com/openai/v1
transcription_api_key: ${DEVCLAW_GROQ_API_KEY}
Store the separate API key in the vault:
vault_save groq_api_key gsk_xxxx
Quick Setup Examples
Z.AI Full Stack (Vision + Audio)
media:
vision_enabled: true
vision_model: glm-4.6v
vision_detail: auto
transcription_enabled: true
transcription_model: glm-asr-2512
transcription_base_url: https://api.z.ai/api/paas/v4
OpenAI Full Stack
media:
vision_enabled: true
vision_model: gpt-4o
vision_detail: auto
transcription_enabled: true
transcription_model: gpt-4o-transcribe
transcription_base_url: https://api.openai.com/v1
Budget Setup (Cheap Vision + Fast Audio)
media:
vision_enabled: true
vision_model: gpt-4o-mini
vision_detail: low
transcription_enabled: true
transcription_model: whisper-large-v3-turbo
transcription_base_url: https://api.groq.com/openai/v1
transcription_api_key: ${DEVCLAW_GROQ_API_KEY}