Running an always-on agent and, separately, a model — what each one needs in memory, what keeps it alive when nobody is logged in, and what the second one really costs once the token invoice is gone.
An always-on agent and a local model are different workloads whose memory floors differ by roughly a factor of eight, and both numbers are published. What the agent projects say about their own footprint, how to compute a model's floor from its file size, and the order to size things in so you do not buy the machine twice.
Oct 30, 2026A rented server is not an always-on agent. Three settings stand between "it runs while I am logged in" and "it answers at three in the morning", and each one is documented by the projects that need it — plus the reboot test that catches the ones you missed.
Oct 30, 2026"No token fees" is a pricing headline, not a cost. Running the model yourself moves spend from a per-token invoice to four other lines — capacity, idle time, context you forgot to count, and your own hours — and only one of them appears on a receipt.
Oct 30, 2026Muse is a personal agent on a machine you never see. Its terms settle what you own — and the rest is four ordinary pieces: a host, a channel, a scheduler and memory on disk. Each of them is documented by open projects, including the one setting that does not survive the rebuild unchanged.
Oct 30, 2026