SM/SWA · The machine room · Relaymark

One door to every model.
Expiring quota goes first.

Every agent on the floor needs a model, and every model sits behind a plan with a clock on it. Relaymark is the private gateway they all run on: one address for Codex, Claude and GLM, the plan that resets soonest is spent first, and every token is counted — including the ones that never went through the door.

Macs six, one key eachBehind one door Codex, Claude, GLMQuotas checked every 5 minutes

Who comes to the door

Codex, CLI and appOn six Macs
Claude CodeEvery new session
The factory's pipelinesAgents working around the clock
A key per MacRevocable at any time

What the floor gets

The plan that resets soonestSpent first, 5% kept in reserve
Sessions that stay putFive hours on one account, cache kept
Every token countedThrough the door or around it, once
A warning before it runs dryPace forecast per window
Nothing expires unseenWhat was left at each reset

How it works

Five rules for the power room

Subscription limits come in windows that reset on their own clocks. Relaymark's whole job is to spend them in the right order — and to show what's left.

  1. 01

    Door

    Every Mac, agent and pipeline talks to one private address inside the tailnet, with its own key. Nothing is reachable from the internet.

  2. 02

    Route

    New sessions go to the plan whose weekly window resets soonest, so quota that would expire is used first. Below 5% left, a plan steps back.

  3. 03

    Hold

    A running session keeps its account for five hours, so its prompt cache stays warm. Sessions move only when the preferred plan changes.

  4. 04

    Count

    Requests through the door are counted at the door. Sessions that went around it are read from the logs on each Mac — every token once.

  5. 05

    Warn

    Every window shows what's left, when it resets and whether the current pace lasts. At each reset, whatever went unused is written down.

Why

Why a power room

A small white paper robot watches a row of three gauge dials with needles at different levels, a paper hourglass beside them
Three gauges, three clocks: the one that runs out first gets used first.

Quota is a clock

A subscription limit isn't a pile, it's a timer. The skill is spending the one that runs out first — before its window closes with quota still in it.

Caches like to stay

Switching accounts in the middle of a session throws its prompt cache away. So sessions stay where they started — until there's a reason to move.

Private by design

One owner, the owner's own plans and machines, inside a private network. Not a service, nothing to sign up for. Built on the open-source CLIProxyAPI.

The power room

Every agent on the floor runs on it.

Relaymark picks the plan for every session. The floor just keeps working.