Base-URL swap
Point your OpenAI or Anthropic client at the Firewall's URL instead of the provider's. Same SDK, same request and response shape — nothing else in your code changes.
Planned · in development · join the waitlist
OpenAI's budget alerts don't stop the bill. We do.
"…project budgets do not enforce a hard cap on spending."
Spend Firewall enforces a hard cap that actually blocks the next request — configured per key to fail-open (stay available, uncapped) or fail-closed (block until healthy again). Your choice, never hidden.
01 · The problem
Budget alerts tell you after the money's already gone.
Every major provider ships the same mitigation: an email or dashboard warning once you cross a threshold. None of them stop the request that crosses it.
Retry loops, runaway agent tool calls, and misconfigured batch jobs are a documented failure pattern in production LLM systems — not a hypothetical edge case. The standard mitigation is an alert that arrives once the spend has already happened, not a system that stops it from happening.
02 · How it works
Point your SDK at one URL. Nothing else changes.
The enforcement lives in the request path, not in a dashboard you have to remember to check.
Per-key spend caps
Set a day, week, and month limit per API key. Whichever cap trips first blocks the request — no averaging across keys, no shared pool to hide behind.
BYOK
Bring your own provider keys. The Firewall proxies to your own OpenAI or Anthropic account — it never becomes a reseller and never sees a key you didn't give it.
Streaming — planned
Token-by-token responses metered and capped mid-stream. The hardest part of proxy enforcement, and the reason it is scoped as planned rather than promised for launch day.
03 · Not swept under the rug
Fail-open or fail-closed — a per-key choice, sold as a feature.
A hard cap that blocks every request once it's reached only means something once you've also decided what happens when the Firewall itself can't reach its own state — a deploy, a timeout, a network blip.
Fail-open
Requests keep flowing, uncapped, if the Firewall can't be reached — a Firewall hiccup doesn't take your app down with it.
Fail-closed
Requests block until the Firewall is healthy again, so the cap holds even at the cost of availability.
You pick, per key — we don't pick for you, and we don't pretend the choice doesn't exist.
04 · Free tier, real enforcement
The free tier is designed to actually block.
Some proxy gateways gate real enforcement behind a paid plan. Portkey's free tier, for example, does not block past the cap — you're billed, then alerted, the same outcome as going straight to the provider.
The planned Spend Firewall free tier is designed the other way: enforcement from the first key, no card required to see a request actually get blocked. Tier shape only, for now — no prices are set.
05 · Land here, expand into Signals
Every request through the Firewall is a trace, too.
The Firewall doesn't only gate spend — proxied traffic already carries what Signals needs: model, tokens, latency, and cost, per key. Once Signals is live, that traffic feeds it with no separate integration to wire up.