Skip to main content

Solutions · Agent spend controls

Agents that can't
outspend their plan.

Budgets per customer, per task, and per call, checked inside your app before the model runs. The agent asks first, and Limitr answers with what it can afford.

Start freeBook a demo
task_481 · ada@acme
› Summarize Q3 churn for the board
task budget$4.28 / $5.00
  1. ✓plan20k tokens$0.90
  2. ✓search35k tokens$1.58
  3. ✓read40k tokens$1.80
  4. ✕write30k tokensreserve denied · $0.73 left
agent Out of task budget. Asking Ada to approve $2.00 more, or finishing with what it has.
acme · month$4.28 / $500
Steps and amounts from engine 0.8.5 · made-up task

The problem

One loop can burn
a month of margin.

Agents call models, tools, and other agents in loops nobody wrote by hand. A runaway plan spends in minutes what the customer pays in a month.

Rate limits count requests. Agents need budgets that count money.

Customer plan
$30
a month, on Pro
One agent run
212
model calls in 14 minutes
Model spend
$38.60
for that run alone
requests: under the limit■ dollars: no limit at all

Agents as customers

Put your agents
on a plan.

In Limitr, an agent is a customer like any other. Give it a plan with its own limits and budget, link it to the workspace it works for, and share a pool across every agent there. Your internal agents work the same way, on your budget.

Limitr Cloud · CustomersEvery customer type
workspaceagentuser
CustomerTypePlan
research-agentagent
llm412k tokens / 500k tokens
its own, resets monthly
web_search88 searches / 200 searches
its own
workspace_tokens610k tokens / 800k tokens
shared with every agent in Acme
budget$3.10 / $10.00
a day
ensureCustomer('research-agent', 'agent', 'agent', 'Research agent', ['acme'])
Limitr Cloud customers · made-up customers

How it works

Set budgets in Cloud,
enforced in your app.

Put a monthly budget on the org and smaller ones on each team in Limitr Cloud. Your agent opens its own budget for each task. Every step counts against all of them at once, so steps running in parallel can't overspend.

Customers · orgAcme Corp
Team
BudgetSpentResetsSet in
monthlyAcme Corp · shared$312.40 / $500monthlyCloud
research_teamResearch · shared$96.10 / $150monthlyCloud
task_481ada@acme$4.28 / $5ends in 1 houryour app
Whichever budget runs out first stops the next step, before the model runs.
Limitr Cloud customer page · made-up customer
How the engine works →

Budgets

A budget at
every level.

Stack them however your customers buy. Whichever runs out first stops the call, and the agent hears about it before it spends, not after.

Per customer

A shared budget on the org or workspace covers every team, user, and agent under it, with smaller budgets stacked inside.

Per task

A budget that ends with the task. Measure the margin of one run to see what a workflow really costs you.

Per call

Each step holds what it's predicted to use before it runs, so a long answer can't slip past the limit.

Agent-aware

Let the agent
see its budget.

Capabilities in the policy become Claude tools in one call. The agent can check what it has left and plan around it, instead of failing halfway through.

Private networks for your agents →
agent.tsClaude tools
const res = await anthropic.messages.create({
  model, max_tokens, messages,
  tools: await limitr.claudeTools('ada'),   // from the policy
});


for (const block of res.content) {
  if (block.type === 'tool_use') {
    results.push(await limitr.claudeToolUse(block, 'ada'));
  }
}

Give every agent
a budget.

npm i @formata/limitr
Start freeBook a demo