What is cc-visionrouter?

cc-VisionRouter is a transparent proxy that sits between Claude Code and an upstream AI API. Its purpose is to solve a specific problem: when you use Claude Code with a model that does not support images, and an image appears anywhere in the conversation context, the API call fails and the session ends. For long-running autonomous tasks that may have accumulated dozens of turns, this can mean losing a significant amount of work with no way to resume automatically. The proxy inspects every outgoing request. If the request contains an image or a document such as a PDF treated as visual input, it redirects that request to a multimodal model you specify. All other plain-text requests go to your preferred primary model. The switch happens automatically, with no manual model-switching commands needed during a session. It also handles the background requests that Claude Code generates internally, such as compact and summarization calls that run across the full conversation history. Because those run over the entire context, which may contain earlier images, the proxy routes them to the multimodal model as well. Installation is through npm. After installation, running cc-visionrouter start rewrites the ANTHROPIC_BASE_URL setting in Claude Code's configuration file to point to the local proxy and backs up the original value. Running cc-visionrouter stop restores the original settings. Claude Code must be fully restarted after starting or stopping the proxy to pick up the changed setting. The proxy supports the Anthropic Messages API format only and does not support OpenAI-format endpoints directly. An interactive configuration menu walks through model selection, and the setup process runs a minimal test call to verify the configuration before saving it.