Models and context

mecha runs the same agent loop over any model: a hosted one through Anthropic's API, or an open-weight one on your own machine through an OpenAI-compatible server. The loop never learns which provider is behind it. Whichever you use, the conversation has to fit the model's context window.

Page	Covers
Providers	Configuring a provider, prompt caching, and which failures are retried.
Serving a local model	Running llama-server: slots, what -c divides, and the numbers that must agree with your config.
Switching models	One router, several models: loading one is the choice, every surface follows it, and a switch waits for the runs in progress.
Compaction	What happens when a conversation outgrows the window, and what survives it.