Every model fills up or dies. Have a second one ready.
Providers hit a wall. Accounts reset. APIs take a nap. Ordered fallback is how Clerk finishes the 9:00 run instead of sitting in In flight until you notice.
Sep 6, 2026 · 8 min · Jimmy Harika
Every provider fills up or dies.
Not as a theory. As a Tuesday. The bar goes red. The account resets on the 1st. The API takes a nap during your 9:00 invoice run. If Clerk stops because someone else's cloud had a feeling, you did not hire a teammate. You rented a demo.
Ordered fallback is the grown-up version. First this account. Then the other ChatGPT. Then Claude. Then the OSS model on the desk. Mid-task. Not "try again tomorrow."
Octopus already says this on the landing. Stack several accounts for more capacity. Work continues mid-task on a fallback. I am repeating it because people still build a single-key shrine and act surprised.
One key is a single point of failure with a chat UI
You would not put the whole studio on one intern's laptop.
Do not put the whole studio on one API key.
Mail at 7:00, Clerk at 9:00, Maker on Monday, Watch on the pages, Scout on the list. That is a morning. That is also a way to burn a quota before coffee if they share a pipe. Separate keys when you can. Stack two of the same plan when you cannot. The team simply has more room to work.
BYO is not a religion. Bring ChatGPT, Claude, Gemini, Grok, Copilot if you already pay. Or skip BYO and use the OSS models in the $99. Same desk. Fallback is the rule in both worlds. Open models go down too. Your laptop goes down too. Have a second path.
Ordered, not "whatever is cheap today"
Cheap-today is how you get a refund letter that sounds like a different company.
Write the order. Clerk's invoices: the model that follows the format skill, then the backup that is close enough, then stop in Needs you if both are drunk. Maker's public copy: the writer you actually like, then a backup you have seen on brand, not a random one that says "delve."
A different model per teammate is allowed. It is the point. Put drafting on the one that writes and analysis on the one that reasons. Do not make Watch write the newsletter because it had leftover tokens.
When you fail over mid-task, the workspace has to be the memory, not the model's mood. Chat memory dies. The brief should not. If the second model cannot see the skill, the footer, and last Thursday's decision, failover is just a new intern.
Stack two accounts before you need them
Add the second ChatGPT on a calm day.
Add Claude before the 9:00 run, not during it.
Test the order with a boring job. A weekly report. A page Watch already knows. Read the receipts. Who, which model, when. If you cannot see the failover in the log, you do not have failover. You have hope.
Caps still matter. A fallback that is "unlimited" in your head will become unlimited in production. Daily caps. Scopes. Vault for every key, not a sticky note. The second account is still a secret.
What you should not do
Do not pretend outages never happen. They do. They will happen on a Monday.
Do not put the backup on a model you have never read. Failover to a stranger is how the Hartley refund grows a personality.
Do not share one login across five teammates and call it stacking. Stacking is capacity. Sharing is a blur. Receipts get worse. Firing one teammate gets worse.
Do not retry forever. If both models fail, park it in Needs you. You taking the tap is cheaper than a loop that spends the month arguing with a 500.
Costs, without a fairy tale
There are no per-task AI charges from us. BYO work runs on your plan. OSS work is in the $99.
That does not make tokens free on your side. A stampede of scheduled jobs will still eat a ChatGPT cap. Stagger the clock. Put a cheap model on the obvious pulls. Put the expensive one on the judgment. Fallback is also a cost tool. When the good model is gone, the next one should still be allowed to finish the list, not the poem.
Seven-day money-back. Dedicated workspace, never train. Export if you cancel. None of that helps if your only key died at 9:01 and the overdue reminders never moved.
The boring test
Unplug the first model in your head.
Does the 9:00 run still finish?
If yes, you have a team.
If no, you have a vendor. Add the second account. Set the order. Keep the brief on the desk so the second brain can read. Then go to lunch like a person who planned for Tuesday.
Straight answers
- What happens when a plan runs out of usage?
- Add a second account of the same plan, or set another provider as a fallback. The job carries on mid-task instead of waiting for a reset. That is the point of stacking.
- Should every teammate use the same model?
- No. A different model per teammate. Drafting on the one that writes. Analysis on the one that reasons. A cheap OSS model on the obvious stuff. Judgment on the one you trust.
- Do I have to bring my own keys?
- No. $99 includes OSS models, the computer, and the board. Bring ChatGPT, Claude, or Grok if you already pay. Same desk either way. Fallback still matters on our pipe and on yours.
Hire your first AI teammate
Ten minutes to set up, on the AI plan already sitting in your browser — or on ours.
Sign up