Docs

How budgets, keys, rules and suggestions work.

Building your own dashboard? Read your spend from your own code with the API.

What Spend does

Spend keeps a budget on what your AI calls cost. Before a call runs, the platform works out roughly what it will cost and checks it against your budgets and rules. Under the warning level it goes ahead. Between the warning level and the limit it goes ahead and you're warned. At the limit it is stopped and nothing is spent.

If the check itself can't be made, the call is stopped too. A budget you can't rely on isn't a budget.

Spend is included with programmable.run. Some parts are still being switched on across the platform; the page for each one says plainly when it isn't working yet.

Budgets

A budget is a dollar amount per day or per month, for everything you spend or for one agent, workflow, model or app. Days and months are in UTC. The smallest budget is $0.05, because each call is checked against an estimate of about 2 cents first.

Lowering a budget, or adding one, takes effect at once. Raising or removing one asks you to confirm first, and every change is kept in the history on Budgets, with who made it.

New here? Start with a template (Careful, Steady or Busy). A template only adds budgets you don't have yet; it never changes one you set.

A call is checked against an estimate before it runs and counted at its real cost after, so one expensive call can end slightly over a limit. The next one is stopped.

The kill switch

Pause all AI spending from Overview with one button. Every call on your account is stopped before it runs, whatever your budgets say, until you resume. Pausing takes effect at once; resuming asks you to confirm.

Warnings and alerts

When a budget reaches its warning level (80% unless you change it), you get one email, once per day or month for that budget. A day with unusually high spending, well above your usual, gets one too.

In Settings you can turn warning emails off, get a weekly summary on Mondays, and send warnings to a Slack channel. The Slack address is kept in the Programmable vault and never shown again.

You can also have each warning POSTed as JSON to your own https address (a webhook). Every request carries an X-Programmable-Signature header: t=<unix seconds>,v1=<hex>, where the hex is HMAC-SHA256 of "<t>." followed by the raw body, keyed with the signing secret shown once when you set the address. Check it in constant time, refuse timestamps more than 5 minutes off, and use X-Programmable-Delivery to ignore a repeat.

Keys and who pays

A call is paid with your own provider key or with the platform's keys and your Programmable credits. Your provider keys stay in the Programmable vault; Spend never stores or shows them.

Each programmable.run key can have a monthly dollar limit, with a warning first. Make a test key on Keys and agents to try an integration safely: $1 a month, 50 calls a day, and it stops working after 7 days.

Routing rules and suggestions

A routing rule sends matching calls to another model, or stops them. A safety rule always wins and may stop a call. A savings rule only switches a call to something cheaper, never stops one, and is skipped if the switch would cost more. Changing a rule from one kind to the other asks you to confirm.

Before adding a rule, try it on your last 30 days of calls to see what it would have switched, stopped and saved. Overview shows what your rules saved this month.

Once a week the platform looks at your calls and suggests cheaper routes, and budgets that fit how you spend. Nothing changes until you switch one on.

Calls and records

Calls lists each AI call with its app, model, who paid, tokens, result and cost, and what finished task it touched. For each call we keep its model, size, cost, whether it worked and who it was for. We don't keep prompts or replies.

Download your daily totals or each call for any month as a spreadsheet file from Calls.

Teams

In a team workspace, owners and admins set budgets for the whole team and for each person. A member can't raise a budget an admin set, and sees only their own limit. Team shows spend by person and a monthly statement you can download.

Connecting your own apps

Calls made through programmable.run and the platform's tools are already covered.

For your own agents: point any OpenAI or Anthropic app at the Programmable endpoint by changing its base URL (Connect shows it), using a programmable.run key. The endpoint takes API keys only, never a sign-in token. There's also a small SDK and a proxy you run beside your app; both use your own provider key, never one we hand out for you to paste.

Reading your spend from your own code

Your spend, budgets, warnings, calls and finished work can be read from your own dashboard or report with a programmable.run key, never your personal account key. The methods, with examples, are on the API page (/docs/api). It only reads; nothing there changes your budgets.

When a call is stopped

Every reason a call can be stopped is explained, with what to do, on the page "Why a call was stopped" (/refused). Failed calls on Calls link straight to it. If you think the platform itself is down, the Status page (/status) shows whether budget checks are answering.