What is deepseek-vision?

DeepSeek is an AI model provider whose official API only processes text. That means if you want to send an image to a DeepSeek model, or have it search the web, it cannot do those things on its own. This project is a proxy service that sits between your AI tools and the DeepSeek API, quietly adding the missing capabilities so your tools think they are talking to a fully featured model. The proxy adds three things. First, it adds vision support: when a request includes an image, the proxy sends that image to a separate vision model (by default Alibaba Cloud's Qwen vision model), converts it into a text description, and replaces the image block with that description before forwarding to DeepSeek. Second, it adds web search: when a tool in your conversation calls for a web search, the proxy intercepts it, queries Tavily or Brave, and injects the results back into the context. Third, it adds web fetch, which lets the model read content from a URL, with built-in protections against certain types of server-side abuse. The proxy also speaks two different API dialects. Tools built for Anthropic's API format and tools built for OpenAI's API format can both connect to the same proxy. The proxy translates between those formats and routes everything through to DeepSeek, so you can use a single DeepSeek API key with a wide range of AI tools without changing how those tools are configured. Setting it up requires Docker or a local Python environment. You provide a few environment variables: a master API key that clients use to authenticate with the proxy, your DeepSeek API key, and optionally keys for the vision and search providers. A built-in web interface at the service's root URL lets you update configuration and restart the service without editing files directly. The project is released under the MIT license and notes that it is a companion open-source project to a hosted agent service called Proma.