Introducing the Recovea Baseline

Today, we're introducing the Recovea Baseline, a free seven-day report of what your own AI calls cost. The package installs beside the provider client you already use, and seven days of reporting later the page holds your own numbers: by agent or workload if you run your own, by client or customer if they pay for what your AI does. When the Baseline closes, a computed Assessment appears in the same place, also free.

The Baseline is available today for applications in Node and Python. There is no card and no charge. It runs for seven full days of reporting and then closes, and the page counts the days, never the calls.

Your calls reach your provider exactly as before. Your own application reports the envelope of each call after the call has returned, so nothing of ours is on your path; if our endpoint were down, your call would not know. The envelope is fifteen typed fields—provider, model, token counts, status, latency and your tag—and never a prompt, a completion or a provider key. Every figure on the page says whether it is reported, modeled or declared, and there is no savings figure anywhere in it.

What the seven days show

The page fills as events arrive. Each section is computed from the envelopes your own application reported, and nothing else.

• The headline. Your reported spend for the seven days and what it annualizes to, with the gap against your last invoice total if you declare one.
• The worst hour. The most expensive hour in the seven days, as a multiple of your median hour.
• Attribution by unit. How the week's calls split across the units you tagged: the distinct units, and the busiest ranked by reported calls with each one's share.
• Where it is concentrated. The models and units that carry most of the reported calls, ranked.
• Loop and retry signature. Runs of near-identical calls at a fixed cadence, retry storms and fan-out spikes, found in the envelopes and never in what the calls said.
• Cost per request, with no success signal. On this wire the absence is the finding.
• The exposure ceiling. A modeled bound: what one runaway loop at your own peak rate would cost over twenty-four hours, beside whatever bounds it today.
• Three closed doors. Your per-client or per-agent spend breakdown with its largest row already open, exposure by agent or workload, and the five moves ranked as bounds.

A modeled cost is your reported tokens priced from a frozen reference price list, with the list's version printed beside the figure. Export the events and check the arithmetic.

The Assessment at the close

On the day your Baseline closes, the Assessment appears in the same place, at no charge. It is computed, all of it, from the events your own application reported during the seven days. Two sections lead depending on the answer you gave at install: AI Cost Defensibility for firms that bill someone else for work agents did, and Agent Exposure for teams running agents with nobody watching.

• Every section is open at no charge, including sections we add later.
• The five moves, in order, each stating the cost or the exposure it bounds. Never a saving.
• What we could not see, stated as plainly as what we could: the calls with no reference price, the success signal this wire does not carry.
• How to check the arithmetic: the export of every event, the chain head printed beside it, and a published tool that re-derives the chain offline.
• A reliance disclaimer, because every sentence on it is computed and nothing on it is audited. Nobody at Recovea writes it and nobody signs it; a number you can re-derive is the stronger thing.

The Assessment is yours whether or not you go on. Nothing is erased for leaving it where it is, and there is no deadline.

Availability and pricing

Anyone can start a Baseline today. The tap stops reporting on the day the Baseline closes, and the page says so.

Under the Assessment there is one offer: the plan whose band your own figures fall in, at that plan's published monthly or annual price. Where your figures fall under the Free plan's own cap ceiling, the page says so and points at Free. We publish the plans and their prices at recovea.ai/pricing; a plan's price is its band's published price and never depends on what the Assessment found.

Start a plan and reporting resumes on your next call, with nothing to reinstall. The Assessment then re-baselines every calendar month, and the caps the moves point at can be armed, which is the one thing that routes traffic through Recovea. A plan started from the Assessment page begins the day you start it, because the free week on your own traffic was the trial.

Get started

Start your Baseline, say whether you bill clients for AI work, and type the label that fits: a client or matter id, or an agent or workload name. The page shows your key once. Ten minutes, three steps:

1. Install the package. npm i recovea or pip install recovea.
2. Put your key in the environment. export RECOVEA_TAP_KEY="your key"
3. Wrap the client you already use.

const openai = tap(new OpenAI(), { tag: "matter-4821" }); // use openai exactly as before ```

```python from openai import OpenAI from recovea import tap

client = tap(OpenAI(), tag="matter-4821") ```

The tag is the one field you type. Use the real client or matter id, or the agent or workload name; it is the unit the page splits your calls by.

To look before anything leaves, set `RECOVEA_TAP_INSPECT=1` and run your application. Every batch prints on your own machine as the exact JSON body the endpoint would receive, and nothing is transmitted. No key is needed for that, so you can do it before step 2. The fifteen fields are published at [recovea.ai/docs/passive-tap](/docs/passive-tap); if a field is not in that table, it is not in the object.

Start your Baseline at [recovea.ai/baseline](/baseline). What the Assessment holds, and its terms, are at [recovea.ai/assessment](/assessment).