You've got a cloud model, a local model, an agent or two — all running in the background. Quick: what did they cost you today? Nobody knows. TokenBrake shows every AI on your machine in one menu-bar glance — and hard-stops a runaway API bill before it happens.
A cap is a lagging indicator. It does nothing until the money is already gone. If an agent gets stuck in a loop overnight, the cap sits there while the loop eats the entire month's ceiling — and fires only once it's spent. That isn't the failure anyone's actually afraid of.
v2 watches the shape of the traffic instead. One stuck agent, 02:00 to 08:00, nobody awake:
| Budget cap alone | Runaway breaker | |
|---|---|---|
| What the night cost | $280.01 | $3.60 |
| Calls it let through | 23,334 | 300 |
| Time to act | 5.8 hours | 10 seconds |
Same cap, same prices, same loop. The only difference is that v2 recognised what it was looking at: “100% of 12 calls in the last 10s were the same request, at 73 calls/min. That is a stuck loop, not work.”
A loop running the same request over and over — including the nastier version where the conversation keeps growing but the question never changes. Plus raw dollars-per-minute, retry storms, and a surge far above what that agent normally does.
A fast batch job with 200 different prompts sails straight through. High throughput isn't a runaway — repetition is. A false trip costs more trust than a missed loop costs money, so it's built to be quiet.
A real circuit breaker: after the cooldown it lets one call through to see if the agent recovered. Recovered, it closes. Still stuck, it backs off. A false trip costs one cooldown — not a 3am phone call.
Every path is wrapped. If detection itself throws, your call goes through. TokenBrake is a brake, not a kill switch on your business — the only time it blocks is a trip it can explain to you in a sentence.
There are two kinds of AI cost, and no tool on earth puts both in front of you.
GPT, Claude, Gemini, Grok, Groq, DeepSeek, Mistral and OpenRouter — billed per token. A runaway loop or a leaked key turns a $10 month into a $2,000 one overnight. TokenBrake meters them all live and can stop them cold at your budget.
Ollama and other local models cost $0 in API — but real electricity, RAM, and disk, quietly, 24/7. TokenBrake reads your machine and shows you the number nobody else measures.
Set a budget. Pick a mode. That's the whole product.
The circuit breaker. Hit your budget and the next call is stopped with a clear message — not a surprise invoice. Free for almost everyone; $249 once for a company over the threshold.
For when the spend is a profit center and you want to run over. TokenBrake just warns you and keeps going — you always know exactly where you stand, and you stay in control.
Install the widget. Grab SwiftBar (free menu-bar host) and point it at TokenBrake. Your local models show up automatically — no setup, nothing to configure.
Meter your cloud AIs. Change one line — the API base URL — to localhost:8787/openai. Your key passes straight through and is never stored.
Set a budget. set-budget openai 20 hard and you're protected. Live spend shows in your menu bar; a runaway bill gets stopped before it lands.
No plans, no trial clock, no per-seat maths. The full safety brake — unlimited agents, every cap, all of it — is free for individuals and for any company under 100 people and under $1M revenue. Companies above that buy one licence, once.
Because a solo developer capping a $40 bill and a company running twenty agents burning thousands are not the same customer, and only one has a budget. The people who can't afford it don't pay. The people it saves thousands pay $249, once, forever.
The code is public and readable — read every line before you run it in front of your production traffic. The licence is PolyForm Small Business, not MIT. It was MIT for two weeks in July 2026, and changed before anyone had adopted it — that MIT release stays MIT forever and we can't revoke it. LICENSE-HISTORY.md in the download says so plainly, and so does this sentence.
What counts as an "agent"? One labeled stream of AI usage you route through TokenBrake — a bot, a script, a project. There's no limit now — monitor as many as you like.
We're a meter your AI traffic passes through — so we're honest about the line.
127.0.0.1, so it meters what runs on that box. That is a security decision, not an oversight: there is no auth to misconfigure and nothing exposed to your network. Run one per machine.Bottom line: this is for the API-based AI your business runs — the agents and scripts doing the work. It is not a way to track an employee's personal use of a chat app.
We sit between your app and your AI provider. That's a position of trust, and we treat it that way.
TokenBrake is a safety tool, not a warranty — it dramatically reduces bill-shock, it doesn't promise a bill can never exceed a budget.