Spend credits well
Usage Last 30 days
AI usage is metered in credits, bought outright rather than subscribed to, shared by the whole workspace, and they do not expire. Nothing renews; whatever you do not spend stays on the balance. That is the entire model, and it means the only question worth asking is where the money actually goes.
This guide is about that. Current prices live on /pricing, and deliberately nowhere else.
Look before you tune
Settings → Usage is the page. Day-by-day spend for the last 30 days, a breakdown by feature - chat, tasks, research, routines - and headline figures including how much of the input was served from cache and what that caching saved. Every row expands into who spent it and which models ran.
Almost every team that thinks it has a credits problem is surprised by that page. The expensive thing is rarely the thing anyone noticed. Spend five minutes there before changing a single setting.
The two big spenders
Coding runs and app builds are, by a distance, the heaviest things a workspace does. Both are also the two you can move off credits entirely: a run started from the Mac desktop app uses the coding agent already installed on that machine - Claude Code, Codex, Grok or Hermes - on that person’s own subscription, for no credits at all. Engineers who already pay for one of those are usually delighted to.
Settings → Agents → Spend can make that the rule rather than the habit: credit runs can be switched off separately for agent tasks and for app builds. With one off, that surface no longer runs on WeMachines at all, and a browser is told plainly that it cannot start one. Everything else - chat, routines, voice, summaries - carries on.
Spend
The ceiling nobody regrets setting
The model picker offers the whole catalog, which is the right default and also means the dearest models in the world are one click away. On the same Spend page, a price ceiling - a highest model price in dollars per million output tokens - narrows every picker in the workspace to what is under it.
It never breaks anything. A model above the ceiling is quietly resolved down to one the team allows rather than refused, so an agent somebody pinned to an expensive model before the ceiling existed keeps answering, more cheaply. Leave it blank and there is no ceiling, which is where every workspace starts.
The quiet one: what agents re-read
By default an agent sees the whole team chat, including every other agent’s earlier messages. That is what stops two agents contradicting each other, and on a busy workspace it is also most of what the AI’s reading costs - agents get asked for long answers, and every one of those is re-read on every later run.
Settings → Context has a switch for it: take AI-written messages out of the transcript and leave it to what people said. It can save a talkative team a great deal. What it never changes: an agent answering in a thread always sees that whole thread, and an agent that needs the wider room can still go and read recent chat itself.
Which docs ride along
Every doc has an AI context setting, on its row in Docs and on the page itself. Title only is where every doc starts: the AI sees the title and opens the doc when a question needs it. Always in context puts the whole doc in front of the AI on every request - right for a roadmap or a style guide, and paid for on every question, including the ones about something else. Hidden from AI closes a doc to the AI altogether; people still read it as before.
AI context
Keep the always-on list short. The context is sent again on every step of a run, so one long doc marked always-on is charged several times per answer, not once. Settings → Context shows what the workspace sends with every request, biggest first, and a doc's setting can be changed right there.
Tasks need no setting, because the board already says which ones matter. An open task rides with its description and comments. A finished or cancelled one rides as its title and status, and the AI opens the rest when a question needs it - so closing work out is also what stops you paying to re-send it.
How long one run may go
An agent works in steps - read something, act, read the next thing - and every step re-reads what the run has gathered so far, which makes the long runs the expensive ones. Turns per run, on the same Spend page, caps the steps one run may take before it writes up what it has: 25 unless you change it, anywhere from 1 to 100. A run that reaches the limit says so, and tagging the agent again picks up where it stopped.
Habits that cost nothing to adopt
- Give routines the smallest context they can work with. A trimmed digest is usually enough; the whole workspace, hourly, is the most expensive line a team can leave running. See routines.
- Clear an app’s thread when it becomes archaeology. A long thread is paid for on every turn.
- Set thinking per agent, not everywhere. An agent that thinks before replying is slower and dearer; a triager that files tickets does not need it, and a reviewer arguing about architecture does.
- Ask once, properly. Three vague mentions cost more than one that says what you want back - and you can tag it again while it is still working, and it will answer both together.
Who may spend, and not running dry
Everyone can see Settings → Billing - the balance, the plan, what has been charged - because what a team is billed is not a secret from the people whose work is being billed. Changing any of it needs the billing role, which an admin on the Max plan grants per person. It is deliberately narrow: a billing member can spend money and do nothing else an ordinary member cannot - it exists because handing somebody the company card used to mean making them an admin, which handed them the whole team as well.
When credits run out, AI replies pause until more are bought - no overage, no surprise invoice, and no silent degradation either. If you would rather not think about it, automatic top-up buys more when the balance falls below a number you set. It is triggered by the balance rather than the calendar, so a quiet month costs nothing.
Two things people assume cost extra and do not: the multiplayer side - connecting your own terminal agents over MCP, the live view of who is working on what, agents moving tickets - and skills. Both are on every plan, the trial included.