BurstyRouter

BurstyRouter sits between your AI tools and the AI models they talk to, deciding where each request should actually go. Its main idea is local first: if you are running an AI model on your own computer, it sends requests there by default, keeping your prompts on your own machine. Only when your local setup is busy, failing, or missing the requested model does it forward, or burst, the request out to a cloud service.

That cloud service is TrustedRouter, a paid gateway the project's creators built specifically to keep overflow traffic private. Rather than trusting a typical intermediary to simply promise it will not store your data, TrustedRouter runs inside a verifiable secure environment, publishes the exact code it is running so anyone can check it, and offers an option where your prompt stays encrypted all the way to the AI model itself, not just during transit.

BurstyRouter speaks the same request formats used by OpenAI and Anthropic's Claude, so existing tools built for those services can point at it without changes. It automatically looks for common local AI servers like Ollama, LM Studio, llama.cpp, and vLLM running on your machine, and you can also rename models, so a tool that expects a specific cloud model name can be quietly redirected to a different local model instead. There is a detailed set of rules controlling exactly when a request stays local, when it is forced to stay local, and when it is allowed to burst to the cloud, including behavior for connection errors, timeouts, and a local queue that is completely full.

Installing it is done through Homebrew, downloading a prebuilt binary, using the Go install command, or building a Docker image from the source code yourself. Settings are controlled through command line flags or matching environment variables, covering things like which port it listens on, how many local requests can run at once, and an optional spending limit for cloud usage.