No spin. Here's exactly what TokenBrake does, what it doesn't, and how it's different from the limits your provider already gives you.
No. TokenBrake runs on your own server. Your provider API key passes straight through to the provider and is never stored, logged, or written anywhere. It keeps only token counts and dollar amounts — never your prompts or responses.
Native limits are account- or workspace-wide and reconcile on a monthly cycle. OpenAI's can lag up to 24 hours; Anthropic's own docs call its spend reading "informational, not transactional"; Gemini's mandatory cap is account-wide and lets up to 10 minutes of overage through. TokenBrake meters every call in real time, per agent, and hard-stops the next call at your budget. See the guides for the details on each.
OpenAI, Anthropic (Claude), xAI (Grok), Groq, DeepSeek, Mistral, Google Gemini, and OpenRouter — including streaming and prompt caching. More on the roadmap; email us to bump one up.
Neither — it's priced by monitored agent. An "agent" is one labeled stream of AI usage you route through TokenBrake (a bot, a script, a project), regardless of the device it runs on. Solo covers up to 3 agents; Business covers unlimited.
Yes for any app or agent whose API base URL you control — laptop, server, or phone. It cannot meter consumer chat apps like the ChatGPT or Claude mobile apps, which bill through the vendor and can't be proxied by anyone. It's for the API-based AI your business runs, not a way to track someone's personal chat-app use.
You pay once a year in USDC on the Base network. Solo is $99/year (up to 3 agents) and Business is $490/year (unlimited). After your payment confirms on-chain, you get your license key instantly on the pricing page.
Nothing breaks. License enforcement is intentionally soft — an expired or over-plan license never stops your traffic, it only prompts you to renew or upgrade. A licensing check must never take down your production.
Yes. TokenBrake serves a machine-readable manifest at /agent.json that any AI agent can read to discover how to check the meter, set budgets, and set alerts. It also ships an MCP server so tools like Claude Desktop and Cursor see TokenBrake as native tools. See the docs.
One line: curl -fsSL https://tokenbrake.com/install.sh | sh. It's a single zero-dependency file that runs on Node 24+ (built-in SQLite) — no database and no Docker required. Docker and from-source options are in the docs.
No. If TokenBrake itself faults, it fails open and lets the call through rather than break your production. It's a safety brake that dramatically reduces bill-shock — not a warranty that a bill can never exceed a budget. Keep your provider-side limits on too.
It runs on your own infrastructure, requires a shared secret (TB_KEY) on every request compared in constant time, only forwards to an allowlist of provider hosts (SSRF-guarded), caps request sizes to resist abuse, and never stores your keys or prompts.
See what your AI is costing you — and what a runaway loop could cost — in the free calculator.
Open the cost calculator →