Introducing Recovea: The Financial Layer of AI

Today Recovea is open to everyone. It is an AI-spend gateway: a thin layer between your code and your model providers that meters every request, enforces your budgets, and writes a receipt for every dollar. The integration is one line: point your client's base_url at api.recovea.ai and keep your own keys. We never resell tokens, and your traffic still goes to the provider you chose, under the key you own.

We built this because AI spend became real money faster than the tools to govern it. Finance sees one number at month end. Engineering sees a dashboard that cannot answer which key, which route, which customer. Nobody holds a receipt. And for a startup, the line this leaks from is not "tooling" -- it is runway, the only budget that cannot be re-forecast. The missing piece is not another optimizer. It is boring, structural financial infrastructure in the request path.

What this protects: runway

An ungoverned AI line spends runway two ways.

Bounded loss. Runway's worst enemy is not steady inefficiency; it is the single unbounded event -- the loop that runs all night, the leaked key, the retry storm. A hard cap converts an unbounded loss into a bounded one. At the cap, the gateway refuses before it forwards, so a refused request is not sent to your provider and carries no provider spend. Alerts warn you on the way there; the cap is the thing that acts. This is insurance arithmetic, and it holds no matter what your month looked like.

Recovered time. Tracing a surprise spike through provider dashboards costs founder-days, and founder time is runway in different units. A receipt for every request -- model, tokens, cost, status -- with spend attributed per key, per project, and per route turns "what were these charges" from an afternoon of archaeology into a filter. Failed requests are a line item too: the money burned on errors is visible instead of folded into the mystery.

The scope, stated plainly: caps act on the traffic you route through the gateway, which is API-key traffic -- a consumer subscription cannot route through any gateway, ours included. And a cap is a stop, not a ceiling: enforcement happens in the request path, but a fast burst can finish roughly one request over the cap before refusals start. Both sentences belong in the pitch, not the fine print.

What is live today

• The gateway. OpenAI-compatible, one base_url change, and designed to fail open -- exactly what that means is spelled out below. Overhead today is ~1ms p50 / 2ms p95 local gateway overhead (measured July 2026, non-production), a benchmark of the gateway hop rather than a production percentile; we will publish production numbers once we have them, not before.
• The meter. Every request metered live across providers (model, tokens, cost, status) while it happens, not at month end -- priced against a frozen reference price list (rpl-2026-08-29). The same table ships inside the MIT-licensed CLI on npm (recovea@0.1.4), so you can check our arithmetic against the bytes rather than against a promise.
• Receipts and the ledger. A per-request receipt and a hash-chained ledger under a published recipe, recovea-chain-v1, that you can re-derive offline. You do not have to trust us to check our math.
• Caps and the kill-switch. Budget alerts at 50 / 80 / 95% tell you; the cap acts. The refusal is an OpenAI-shaped error your SDK parses natively, with a code that says the wall was yours, not the provider's quota. The kill-switch is the same wall thrown by hand, and it has no ceiling on any plan.
• Attribution. Spend broken down per key, per project, and per route, so the bill finally has names on it.
• The free scan. Point it at your billing data and get a teardown of where the money goes, with the method shown alongside the result.

The free tier is capped at 100 requests/min and 200K tokens/min. Paid plans start at 600 requests/min and 1M tokens/min, rising to 3,000 requests/min and 5M tokens/min at the top of the ladder. Rate limits are indicative, never a guarantee: the limiter runs inside each gateway process, so the throughput you actually get can land above or below the published figure. Starting is free and takes no card.

Two failure postures, on purpose

Two sentences on this site can look like a contradiction, and both are load-bearing, so here they are side by side rather than blurred into one.

Availability fails open. If any Recovea-added layer degrades -- metering, caching, budget checks, the ledger -- your traffic goes straight to the provider on passthrough. Your app is designed to keep serving even when we do not: you lose a meter reading, your users still get their response.

Armed caps fail closed. A cap is your own kill-switch, so it is enforced as a money control. When your budget state cannot be confirmed, the gateway refuses the request with a 402 rather than guess; a workspace with no enabled cap is designed to fail open. (Tenant isolation is the one other deliberate exception: an isolation fault refuses rather than serves.)

These are different dials. Fail-open is about your availability. Fail-closed is about your money -- and, isolation aside, it applies only to the wall you asked us to build.

What it costs

The ladder is signed and published end to end, with no talk-to-us cliff at the top: paid plans run from $49/mo to $999/mo, annual billing at -20%, each with a published band of included routed spend. Crossing a band is measured and shown, never billed today -- and if that ever changes, the pricing page changes first. Every paid tier starts as a 14-day free trial, card required, with the conversion date and amount shown at checkout and one-click cancel. Free is not a demo: in-path metering, receipts, the ledger and its export, and real protection. Caps up to $1,000/mo on Free — Developer raises the ceiling to your full band. The kill-switch has no ceiling on any plan.

If you would rather have the finding done for you, there is one service line: the AI Spend Assessment, priced by your monthly AI spend band -- $500, $1,500, or $4,500, quoted above that -- run observe-only on your own live traffic and delivered as a written report, ranked worst offender first. The fee is fixed and never contingent on what we find: our fee cannot want a particular answer. Within 60 days of delivery, all of it credits toward an annual plan.

One sentence is true on every rung: the fee is flat, and we never take a percentage of what you route or of what you save.

The referee takes no money from the players

Recovea is BYO-key. No token markup, no reselling, no provider money -- and never a percentage of your spend or of your savings. If the referee is paid by a player, the score is marketing. The subject of the proof never pays for the verdict. That neutrality is the product, and it is why the ledger is designed to be checked without trusting us.

The ask

Run the free scan on your current bill. It costs nothing and the method is public. If you like what it shows, point one route's base_url at us and watch your real spend go live in under ten minutes, free, no card. The receipts start with the first request; the cap is yours to arm the same afternoon.