Has anyone else run into this problem while building an AI SaaS?

You launch your product, a few customers start using it, and then you realize…

How do you know which tenant is spending your OpenAI/Anthropic/Gemini budget?

And even if you know, how do you stop them before they burn through hundreds of dollars?

I looked around, and most tools focus on observability—they tell you what happened after the API call. I wanted something that could also enforce spending limits before the next request goes out.

So over the past few weeks, I built token-limit, a Python SDK that:

✅ Monkey-patches the official LLM SDKs (OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter)

✅ Automatically tracks every LLM request per tenant

✅ Batches usage events in the background (no changes to your LLM call sites)

✅ Lets you set per-tenant spending limits and raises a LimitExceededException before another paid request is sent

The setup is intentionally simple:

from token_limit import Meter, MeterConfig

meter = Meter(MeterConfig(api_key="..."))
meter.patch_all()

with meter.for_tenant("acme"):
    client.chat.completions.create(...)

Enter fullscreen mode Exit fullscreen mode

I’m still actively improving it, and I’d genuinely love feedback from people building AI products.

A few questions:

• How are you tracking LLM usage today?
• Do you enforce budgets per customer, or just monitor costs?
• Is monkey-patching something you’d be comfortable using in production, or would you prefer another approach?
• Which provider or framework would you want to see supported next?

If you’d like to try it or review the implementation, here’s the repo:

GitHub: https://github.com/AliEzatyar/token-limit

I’d really appreciate any feedback—positive or critical. If there’s something that makes you think, “I wouldn’t use this because…”, I’d especially like to hear it.