what is visionbridge fr?

VisionBridge is a small proxy server that gives text-only AI models the ability to see images. It sits between your chat app and two separate models: a reasoning model that does the thinking, and a vision model that does the actual seeing. When a request includes an image, VisionBridge lets the reasoning model ask the vision model targeted questions about it using tools like look, ocr, scan, crop and compare, instead of the reasoning model needing to understand images itself. The project works with popular local and hosted model backends including LM Studio, Ollama, vLLM, and any OpenAI compatible API, so people can mix and match whatever reasoning model and vision model they already have running. No training or custom model weights are needed. Before the back and forth starts, each image gets a quick one time description from the vision model so the reasoning model has some context right away, and repeated images are cached so they are not processed twice. It can be run quickly with Docker by pointing it at your existing reasoning and vision backends, or set up locally with Python for development. A built in doctor command checks that both backends are reachable and working. Settings such as which models to use, timeouts, retry behavior, and image size limits are all controlled through environment variables, and one instance can serve multiple reasoning and vision model pairs at once under different names. The API mirrors the standard OpenAI chat completions format, so most existing chat tools can connect to it with minimal changes, including support for streaming responses. It also includes safety measures around fetching images from web addresses, blocking requests to private or internal network addresses and capping file sizes, though it does not include built in authentication for public deployments. A trace endpoint lets developers inspect exactly what tools were called and what the vision model reported for any recent request, which is useful for debugging.