I Built an API Proxy That Stops AI Agents From Burning Through My Budget

The $847 Surprise
Last month, my side-project agent hit an infinite loop. By the time I noticed, it had burned through $847 in OpenAI credits. The provider bill came 3 days later. No cap, no kill switch, no warning.
If you ship AI agents, you've probably felt this paranoia. You give an agent an API key and hope it behaves. But hope isn't infrastructure.
What I Wanted

Hard spend caps per key — not "alerts," actual blocks

Pre-flight cost checks — reject the request before it hits the provider

Key isolation — per-agent budgets, not one master key

Multi-provider — one URL, any model, no code changes

No surprise bills — ever
So I built Fuse — a prepaid proxy layer that sits between your agent and the AI provider.
How It Works (3 Lines of Code)
Swap your base_url and key. That's it.

import openai

client = openai.OpenAI(
    base_url="https://getfuse.online/v1",
    api_key="pk_live_your_key_here"
)

response = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[{"role": "user", "content": "Hello!"}]
)

Fuse handles the rest:

Pre-flight estimation — checks cost before forwarding

Spend cap enforcement — hard block at the limit (returns 402 or {"blocked": true, "reason": "cap"} )

Key vault — real keys stored at the edge, never in your DB or logs

Multi-provider routing — OpenAI, Anthropic, Groq, Cohere, xAI, Kimi, DeepSeek, Gemini. One URL.
The Architecture in 30 Seconds

Cloudflare Workers at the edge for sub-50ms routing

Prepaid wallet — top up via Paystack (supports USD, EUR, GBP, NGN, and 10+ others)

Real-time token monitoring — even mid-stream for streaming responses

No SDK required — OpenAI-compatible format, works with LangChain, CrewAI, AutoGPT, etc.
Why Prepaid, Not Post-Paid?
Post-paid billing means you pay for mistakes. Prepaid means you cap mistakes. If your agent goes rogue, it hits a wall — not your credit card.
Pricing

Free: 0/mo,10 starter credit, 5% markup, 1 key

Pro: $19/mo, 0% markup, unlimited keys, webhooks, priority routing

Enterprise: Custom, SSO, audit logs, dedicated workers
What's Next?
I'm shipping Fuse publicly today. If you're building AI agents and sleep better knowing they can't bankrupt you, I'd love your feedback.
Try it: getfuse.online — $10 free credit, no card required.

What do you think? Would you trust a proxy with your API keys? What features are missing for your use case?