Research preview · domestic and overseas tools
One review layer
for many tool bills.
Agent stacks now charge by tokens, searches, image outputs, browser hours, sandbox seconds, characters, credits and seats. Verovo catalogs the billing dimensions, imports the evidence, and separates what the bill says from what usage can prove.
Cataloged ≠ connected ≠ live-validated. The catalog currently has 26 tools; 0 have a live customer-bill validation. No customer savings or recovery claims are made here.
What the product can model
A tool is not “supported” because its logo appears here. Each row records the billing unit, bill source, audit capability and validation state. Unknown providers stay visible instead of being forced into a known format.
| Family | Examples | Bill dimensions | Buyer checks / scoped extensions |
|---|---|---|---|
| Agent platforms | Coze · Zapier · Lindy | credits · tasks · seats · model/plugin pass-through | Estimated credits vs cash bill; third-party cash; included quota |
| Model platforms | Anthropic · OpenAI · Gemini · Bailian · Ark · Hunyuan | tokens · search calls · container/session hours | Token vs tool fee; project/tag allocation; price version |
| Search & web | Tavily · Exa · Firecrawl · Bocha | credits · requests · pages · runs | Depth/source multipliers; included credits; overage |
| Browser & sandbox | Browserbase · E2B | browser-hour · proxy GB · CPU/memory-second | Session boundaries; idle/terminated states; resource specs |
| OCR & voice | AWS Textract · Azure Doc Intelligence · Deepgram · ElevenLabs | pages · minutes · characters | Successful pages; character/minute rounding; layered voice fees |
| Media | Replicate · Coze image/video | output images · seconds · resolution | Request count ≠ output count; failed generation rules |
| Maps & weather | Amap · QWeather · OpenWeather | requests · monthly tiers · quota · minimum charge | Service-family quotas; tier accumulation; free allowance |
Domestic examples: Coze, Bailian, Ark, Aliyun OCR, Baidu OCR, Amap, QWeather, Hunyuan, Bocha. Overseas examples: Anthropic, OpenAI, Gemini, Tavily, Exa, Firecrawl, Browserbase, E2B, Deepgram, ElevenLabs, Replicate, OpenWeather, Zapier, Lindy, AWS Textract, Azure Document Intelligence.
One ledger, two questions
Did the provider charge the agreed amount? Separate invoice facts, platform usage and a rule-based recount. Keep free quota, prepaid packs, credits, tax and unlinked usage visible.
Charge reviewFor a concrete customer dispute, revenue discrepancy or procurement requirement, compare metered usage with billing. Upstream spend is a buyer task. Unlinked usage remains an evidence gap.
Metering integrityEvery source document, row, pricing policy and browser run keeps a hash, locator and audit trail. Missing evidence produces review_required, not zero.
Traceable, not guessedpython3 -m buyer_ledger catalog summary python3 -m buyer_ledger catalog list --region global --role buyer python3 -m buyer_ledger catalog coverage --period 2026-10 python3 -m buyer_ledger report --period 2026-10 --perspective buyer --out buyer.md
CDP attach, not credential handoff
Some bill portals have no API. Buyer Ledger can attach to a user-started Chrome over the local DevTools Protocol, reuse the logged-in session, execute a reviewed recipe, download the bill, and push it through the same ingestion and audit pipeline.
A dedicated profile opens a loopback debugging port. The user signs in and completes MFA; Verovo never handles the password.
HTTPS and allowed hosts are checked before actions. Cross-origin navigation and downloads are blocked before they can enter the ledger.
The downloaded artifact keeps run_id, recipe hash and audit path. Parsing can be deterministic or produce a pending draft for review.
# Local Chrome must already be running with --remote-debugging-port=9222 python3 -m buyer_ledger session cdp --endpoint http://127.0.0.1:9222 python3 -m buyer_ledger fetch --recipe recipe.json --period 2026-10 --cdp http://127.0.0.1:9222 --analyze --out report.md
Security boundary: loopback only; no query tokens; any local process that can reach the debug port can control that browser. Use a dedicated profile, close it after the run, and never point this at a daily browser. Recipe actions are not a sandbox; use platform read-only roles where available.
Research cases, not customer logos
Coze
One enterprise bill can contain credits, model fees, official plugins, third-party plugin cash, voice and storage. “One agent bill” is not one fee.
Anthropic
Web Search is priced per 1K searches and excludes token processing; Code Execution has free hours then container-hour pricing; Managed Agents add session-hour. Four evidence lines for one task.
Amap · QWeather
Quota, prepaid packs, service-family exclusions, monthly tiers and a minimum charge all change the payable amount. A simple “calls × unit price” review is not enough.
These are public-pricing research examples. They do not claim a customer, refund, saving or live integration.
What public billing complaints actually show
A SURE 1.6 study collected 1,949 public discovery records, read 150, and admitted 66 strict evidence records. It is exploratory: audit fails 3/12 gates, Chinese sources were unreachable, and only 3 of 18 required vendors produced admissible evidence.
Users react strongly when paid credits expire, a spending cap fails, auto-recharge is invisible, or the usage dashboard cannot reproduce the invoice. Small amounts can still trigger chargebacks.
Ownership + visibilityModel credits, expiry, auto-recharge, hard caps, included quota and cash billing separately. Keep a queue of unexplained charges rather than forcing them into zero.
Entitlement ledgerCheck whether metered usage is represented in billing. Unpriced or unlinked usage remains an evidence gap; it is not automatically leakage revenue.
Metering integrityNot a complaint rate, not a market validation, and not a claim about any vendor's overall conduct. A larger SURE discovery expansion now holds 50,000 unique public records (HN 49,090); strict evidence remains 66. Discovery volume is not evidence volume. Read the bounded study summary →