Example: solo developer
$95.83/ team / yearEstimated avoided runaway spend, before fees.
- Agents
- 1
- Daily spend / agent
- $20.00
- Incidents / team / month
- 1
- Hours before detection
- 2
Spend less on runaway work
What could you
stop wasting?
A stuck agent can keep making paid calls long after the useful work stops. Leash blocks the next call when it cannot fit your budget. Here is what that could mean for your team.
Start with the example below, then enter your own numbers.
01Stop runaway calls
A hard cap blocks the next request before it reaches the provider.
Try the kill switch →02Send less repeated dataAutomatic tool-output compression trims eligible JSON and repeated logs.
Inspect the exact request →03Start a leaner conversationTake decisions and next steps into a fresh chat with a handoff file.
See how handoffs work →01 / Estimate the cost of a runaway
What if you stopped it earlier?
This calculator covers blocked runaway calls only. It does not add compression or handoff savings to the total.
YOUR ESTIMATE · USD
Runaway spend you could avoid each month
$32.64$391.67 per year if the same pattern continues.
An example from your inputs, before Leash fees. This is not measured customer savings.
One runaway, with your numbers
Without a cap · 4h running$16.67
With a cap · stops by the modelled 5-minute point$0.35
That is $16.32 less per incident × 2 incidents a month.
- Normal team spend / 30 days
- $3,000.00
The assumptions behind the number
- 5× burn rate: the stuck agent spends five times its normal hourly rate. This is a chosen assumption.
- 5-minute stop: we assume the cap runs out then. Your real stop depends on the cap and request cost.
- 24 hours/day, 12 identical months: one agent per incident, with no overlapping incidents or seasonal changes.
The formula
How we calculate it.
D = daily USD per agent · I = team incidents per month · H = hours until manual detection
Monthly gross avoidance = I × (D ÷ 24) × 5 × max(H − 5/60, 0)Annual gross avoidance = monthly gross avoidance × 12Your example: 2 × ($20.00 ÷ 24) × 5 × max(4 − 5/60, 0) = $32.64/month. Calculations use full precision; displayed USD is rounded.
Protection starts with routed traffic. The estimate assumes affected calls pass through Leash and the configured cap is enforced. It counts future calls that could be blocked, not refunds, non-LLM cloud charges or normal productive work. API-equivalent estimates are not invoices.
Per team. per year. under your assumptions.
Try another team size.
Load an example into the calculator. Each uses the same assumptions shown above.
Example: small team
$391.67/ team / yearEstimated avoided runaway spend, before fees.
- Agents
- 5
- Daily spend / agent
- $20.00
- Incidents / team / month
- 2
- Hours before detection
- 4
Example: growing team
$2,770.83/ team / yearEstimated avoided runaway spend, before fees.
- Agents
- 20
- Daily spend / agent
- $35.00
- Incidents / team / month
- 4
- Hours before detection
- 8
02 / Check the enforcement
Does the cap actually stop requests?
We test that separately from the savings estimate. These are the results of a repeatable gateway test.
Measured. repeatable.
40 calls arrived.
35 never reached the provider.
Two gateway instances competed for the same budget. Leash admitted 5 requests, then blocked the next 35 before they could incur provider usage.
- Requests blocked before upstream
- 35 / 40
- Budget overshoot
- $0.00
- Shared budget / accounted usage
- $0.096384 / $0.096384
Controlled integration test · mock LLM provider · 2026-10-08. This measures enforcement, not customer dollar savings.
Download the test result →Reproduce in a Leash checkout: pnpm test:load
03 / Why this matters
Real bills.
Public sources.
These independent reports show what unchecked usage can cost. They are not Leash customer savings. Each card explains what happened and which part Leash could protect.
an unattended agent kept repeating requests.
$500+In Claude Code issue #35166, jialejiahi reported more than $500 in unexpected token usage after requests repeated for hours in an unattended Anthropic API session.
Self-reported cost, not an audited invoice or an Anthropic-confirmed root cause. No Leash involvement or measured savings is claimed.
Source checked 2026-10-08.a static site received a six-figure bill.
$104kA site owner documented a $104k Netlify bandwidth bill in Netlify’s support forum. Netlify’s CEO said the user would not be charged.
The bill was waived. Netlify disputed the DDoS characterization. This was cloud hosting, not an LLM loop; Leash does not control hosting bandwidth.
Source checked 2026-10-08.unexpected requests reached a private bucket.
>$1,300Maciej Pocwierz reported an AWS bill above $1,300 after other systems sent requests to his S3 bucket. He reported that AWS canceled the bill.
AWS subsequently removed charges for unauthorized HTTP 403 requests from outside the owner’s account or organization. This historical cloud case is outside Leash’s LLM proxy coverage.
Source checked 2026-10-08.