Free. Private by construction. No API key, no account, no bill - you supply the compute, so there is nothing to pass on.
typed-local is the free tier of typed. Your turns go to a model server running on your own machine - a llama.cpp llama-server (or the Snapdragon NPU engine) - and never to a hosted API. The prompt never leaves the box.
It is the default tier on a machine with no key: typed --model typed-local works on a fresh install, before you have any account anywhere.
You supply the compute, so there is no upstream bill. No monthly quota, no metering, no surprise.
Prompts and code never leave your machine. Not a policy promise - a property of where the model runs.
No sign-up, no key, no login before your first turn. Install and go.
With the model already downloaded, a session keeps working with no network at all.
typed can size and print the exact server command for your machine. Run:
typed local # prints the llama-server command for this machine
Copy the printed command, run it, then start typed. typed doctor verifies the connection and names the model your server is actually serving.
32 GB of RAM for the bundled preset - a hard floor, not a recommendation. The model weights, the server's runtime overhead and the KV cache have to fit, and roughly 6 GB is held back for your OS. typed local prints the whole accounting with your machine's own figures if it turns you away.
All consequences of running on one machine - and each surfaces explicitly rather than being papered over by quietly re-routing to a paid model:
Want frontier quality or a bigger window? Use typed-openrouter - same agent, your own key.