mrmr
mrmr, pronounced “murmur,” starts with an event instead of a prompt.
say a service fails on a machine you really care about, or a feed you follow returns a new record. in other words, something changes that might deserve attention. most of those events should not wake a large agent or interrupt a person.
mrmr is intended to make that distinction with small decision models, while keeping control in ordinary software.
judgment is not authority
the core rule is simple: the model provides judgment; the runtime owns authority.
events enter through an API, an HTTP poller, or the Linux system journal. the runtime stores them, removes duplicates, applies filters, and asks a model for a structured decision. a configured policy then decides whether to ignore, notify, act, or delegate bounded work.
what works now
the project includes a Go runtime, SQLite storage, support for local d1 and other decision providers, and browser tools for configuration and review.
stored events, decisions, and execution traces make the processing path inspectable. shadow mode records what a policy would do without executing its effects. labels and evaluation tools help compare model decisions against human judgment.
right now, i have it monitoring services on my home lab.