cheap-eyes 0.3.0: no settings file to write anymore

When I shipped the first versions of cheap-eyes, setting it up was honestly awkward. You had to write a JSON config by hand, list the folders, and know which OpenRouter models to pick. Fine for me, a wall for anyone else.

0.3.0 removes that:

• Claude Code: one command, npx -y cheap-eyes init. It checks the network, opens the browser to sign you in to OpenRouter, picks the models and connects the server.
• Claude Desktop: download the extension from the release and double-click it. The settings ask for the key, a folder and, if you need one, a proxy.
• Home Assistant: the add-on creates its own access token and shows the connection address in the log.
• Docker: a ready-made image.

You no longer choose models up front. On the first job the server takes the cheapest suitable ones, runs a short test on each and keeps the ones that pass. One caveat: this finds a model that works, not the one that reads best. If quality matters, there is a benchmark you can run on your own documents.

Why bother at all. On my 887 KB log, the task “find everything about the power supply” cost 8 cents and about 12k tokens of the agent’s context, instead of 227k for reading the file whole. That pays off for reading by meaning. If you are looking for an exact value, grep is faster and more reliable, and cheap-eyes says so.

You need an OpenRouter account with credit on it. OpenRouter is not reachable directly from some countries; then you need a proxy, and the setup will ask for one.

Open source, MIT: ai.lounge.ai/@st412m/cheap-eyes github.com/st412m/cheap-eyes