For teams buying AI services — business, engineering and finance
Know the cost.
Verify the value.
Connect charging rules, spending records and delivery evidence with your business requirements. Know what matches, what needs attention and what remains unknown — before payment, acceptance or renewal.
Connect usage, financial records and business outcomes. Calculate task costs and cost per verified success with the local economics command. Try task economics →
Local tools support selected recounts and explicit charge/delivery criteria. Enterprise review scope and fees are under validation, with no recovery guarantee. See the buyer workflow and limits.
Working on local agent logs instead? Use the free developer tools or the free measurement runs. 中文入口：常见问题（账单核对与免费任务）。
CLI v0.9.0 on GitHub; the PyPI project verovo is pending its first publish (install from the tag meanwhile). SDK v0.3.0. 511 Healthcheck tests verified on 2026-10-30. Customer payment and repeat use remain unverified.
A billing question and a renewal decision need different evidence.
Check a specific charge against the applicable agreement and records. A supported difference can enter vendor review; it is not yet a refund.
Check observable work against your stated business requirements. Commission delivery review independently, even when every charge matches the agreement.
Operations explains the work. Finance or procurement owns the spend. We start by identifying the task, available evidence and who can act.
Missing evidence stays unknown. A stricter business requirement does not automatically make a charge incorrect. The local review compares explicit criteria with supplied observations. Broader customer workflows still need validation; there is no general automated quality score.
Two conclusions. One inspectable review.
The main-preview review command reads your scoped task and local JSON evidence. You supply the applicable charge rules and business criteria; each check shows its observation, rule reference and evidence hash.
The synthetic invoice matches the supplied agreed total. This conclusion applies only to that check.
Charging criterion metThe same example exceeds the buyer’s limit for tasks needing human rescue. Correct charges do not settle acceptance.
Business criterion unmetA missing completion record remains UNPROVABLE. The report names what is missing instead of assuming success or failure.
Unknown stays visiblegit clone https://github.com/roy-tong/verovo.git cd verovo python3 healthcheck/verovo review --sample --html buyer-review.html
Synthetic example, not a customer result. New on main; not included in the v0.9.0 tag. A person confirms rule applicability and source truth. Prepare your own task →
AI support is an initial application scenario. Coding agents, automation platforms, tools/APIs and model services enter through specific buyer tasks; capabilities are verified one scope at a time.
Many tool bills. One evidence layer.
Agent tools charge by tokens, searches, credits, browser hours, sandbox seconds, pages, characters, images and seats. Buyer Ledger catalogs the billing dimensions, imports the evidence, and separates invoice facts from what usage can prove.
Check platform, direct and cloud-marketplace charges against usage and applicable pricing, including free quota, prepaid packs and credits.
Charge reviewReview upstream spend as a buyer. Downstream billing reviews extend from a concrete dispute, revenue discrepancy or procurement requirement.
Transaction-led extensionUse authorized exports, APIs or local files. Browser collection is an optional access method when the task warrants it; portal coverage is not assumed.
Task before integrationpython3 -m buyer_ledger catalog summary python3 -m buyer_ledger catalog coverage --period 2026-10 python3 -m buyer_ledger fetch --recipe bill.recipe.json --period 2026-10 --cdp http://127.0.0.1:9222 --analyze --out report.md
Research preview: 26 tools cataloged across China and overseas. Cataloged is not connected, and 0 tools have live customer-bill validation. See the support matrix and limits →
From records to a decision your team can use.
Name the vendor, period or decision date, available record types and responsible people. No full raw export is needed for the first conversation.
Agree the supported scope. Keep findings in both directions, applicable rules, rerun steps and missing evidence visible.
Separate vendor acceptance, actual refunds or used credits, buyer decisions and any agreed service fee. A report alone is not realized value.
Procurement advisers and support implementers can bring a specific client task. We start with bounded work and clear responsibilities.
Describe a buyer task Try the free billing recountDeveloper tools: inspect your own logs.
For developers using Codex or Claude Code: first try a synthetic example. You need Python 3.9+ and a terminal on a computer. No logs or account required.
# PyPI project `verovo` is pending its first publish; install the 0.9.0 tag for now python -m pip install "git+https://github.com/roy-tong/verovo@v0.9.0#subdirectory=healthcheck" verovo demo
Open verovo-demo-report.html in the folder where you ran the command. The terminal prints the exact path. The example contains a retry chain and failed executions; these are demonstration findings, not installation failures.
Just browsing? See what an explanation looks like
The synthetic demo shows 1 retry chain containing 3 executions, and 7 failed executions. That does not prove seven bugs or seven charges. A useful next check is to inspect the first error in the longest failure run.
Read an illustrative diagnosis. To inspect your own sessions, run the CLI on the computer holding your logs.
Installation help: macOS, Linux, Windows
Use Python 3.9+. On Windows, use py instead of python if needed. On macOS/Linux, use python3 if python is unavailable.
If pip says “externally managed” or lacks permission, use a local virtual environment:
python3 -m venv .venv .venv/bin/python -m pip install --upgrade verovo .venv/bin/verovo demo
Windows PowerShell:
py -m venv .venv .\.venv\Scripts\python.exe -m pip install --upgrade verovo .\.venv\Scripts\verovo.exe demo
No activation required. Use the same executable path for later commands. If verovo is not found, check the install environment rather than reinstalling globally with administrator privileges.
Next: inspect your own logs and prepare a shareable summary
verovo check --json report.json verovo share report.json --out summary.json verovo validate summary.json
Auto-discovery prefers Codex when both runtimes have logs. For Claude Code use verovo check --runtime claude --json report.json. Open the HTML report at the path printed in the terminal; keep the full JSON and HTML local. Inspect summary.json before choosing to send it. Exporting is not uploading.
No logs, an error, or an unexpected result?
No logs in the last seven days? Try verovo check --all, or point to the real directory with --dir "/path/to/logs" --runtime codex (use claude for Claude Code). An empty directory produces an input error, not a clean bill of health. If a command fails, stop before running the export commands.
Healthcheck exit 0 means the report ran, even if it contains findings. Exit 2 means an input/output problem. UNPROVABLE means missing evidence, not failure or a clean result. For conformance, exit 1 means a failed or required-but-unproven invariant.
Report a reproducible tool bug with the version, command, expected behavior and a redacted error. Do not attach raw logs, report.json or credentials to public issues. Reporting a reproducible tool bug is free.
中文使用说明 · Maintainer / metering quickstart
Review coverage and findings. The CLI and public checks stay free.
For a free measurement run, bring a sanitized sample. Enterprise charge or delivery review starts with a separate task-fit discussion.
Free sample checks usually take 48 hours after a clean sample. Enterprise scope, timing and any fee are agreed for the specific task.
Similar events. Different facts.
The distinctions reliable agent metrics depend on — kept separate, each with its own evidence bar.
Attempt≠Operation
Attempts are execution facts. Operations are logical intents. Retries are multiple attempts of one operation, not “2 operations with 50% success.”
Available≠Influential
A tool result appearing in the next context proves availability. Influence — that it changed the agent's behavior — needs counterfactual evidence. Reference is a weak estimator, not a state. Unknown is a valid, first-class answer.
Observed≠Inferred
A selection is not a preference. A model, router, policy, workflow, or user can make it. Verovo records what happened before inferring why.
We audited the AI usage-tool ecosystem.
The same five bugs kept appearing.
One assistant message billed per content block (2–5×). Re-emitted Codex events counted twice. Cache writes priced at 1.25× instead of 2×. Price tables 5× off — including in the industry's pricing source of record. Sessions wiped on resume. Every finding: pinned commit, runnable repro, public filing — upstream responses linked in the report, an external maintainer's 136k-event corpus review challenged an initial finding and led to a correction.
Audit counts are from the September 2026 report; they describe technical findings, not paid adoption.
Read the Token-Accounting Bug Report →
External methodology correction: the author revised 436k to 54,154 tokens; not a dollar invoice →
Maintain a usage tool? Audit it in 10 minutes with our fixtures →
Everything here is free.
Run the tools yourself, or bring a sanitized sample and get a human-run measurement check — free, usually within 48 hours of a clean sample. The $10/$20 developer diagnostics previously listed here were removed on 2026-10-21; these sample runs and the open-source tools remain free. Enterprise buyer reviews have a separate, individually agreed scope.
- Local Healthcheck and conformance
- Open-source SDK and public rules
- No account or upload
- Your 20–100 sanitized rows, our check
- Findings as checkable claims; UNPROVABLE where your surface is blind
- Fixture published under your handle only if you allow
- External measurement runs for MCP/API providers
- Mapping answered in writing; drift fails loudly
- Public thread, so the next provider reads it
If the check was useful, a star on the repo helps others find it — never required, never a condition.
Paying never changes what the evidence says.
Rules we wrote down before anyone asked us to bend them.
The spec, the engine, the SDK, conformance, and the local dashboard stay open. The free list is written into governance — not revocable later.
Tools stay openRaw data never leaves your machine by default. You choose whether to send an aggregate summary or a manually redacted sample for a diagnosis.
Your traces stay yoursNo paid placement, no recommendation slots, no custody of funds. The moment a referee sells shelf space, the measurements stop meaning anything.
Findings stay reviewableThe measurement toolkit
is open source.
The semantics, rules and conformance vectors are public on GitHub. Maintainers and tool builders: this is your entrance.
What ships today
- ● Canonical observations - one portable schema
- ● Operation / attempt correlation, fail-closed
- ● Conformance vectors in CI — PASS / FAIL / UNPROVABLE
- ● Local analytics without any cloud
- ● Deterministic replay - reproducible numbers
Numbers track releases and the roadmap · last verified CI runs.
No agent-side install
Third-party agents never need Verovo. Measurement happens where the capability is delivered.
Provider-side measurement
The SDK wraps your tool handlers and emits canonical observations using one schema and six payload classes.
Not on the critical path
Non-blocking, durable spool with loss accounting. If the SDK fails, your server doesn’t.
Your data can stay local
JSONL on your machine, analytics generated locally. Hosted ingestion is not part of the current offer.
MCP is the first reference surface, not a requirement. Closed-source software is fine.
Our commercial validation starts with a concrete buyer billing or procurement task. Healthcheck and Metering QA remain separate developer services. Contribute rules, counterexamples or adapters to improve the open tools; free adoption is not a sales obligation.
The next task starts with the same buyer.
A billing review may reveal a separate acceptance or renewal question. We explore that work only when the buyer has a real decision, an agreed scope and a reason to commission it.
Establish usable evidence, a review path and an action the customer can take.
Learn whether the same team commissions another period, acceptance review or renewal task.
Expand rules and vendor coverage when customer demand and delivery economics justify the work.
These are validation steps, not established recurring revenue. Calculation reproduction supports the review; full execution replay and training-data selection are not current product commitments.