AI agents give different answers each run, and if they can call tools, a wrong answer can issue a refund or change a record. Hand checks do not scale. CertiQA tests agents before release, and also tests web UIs and APIs.
For engineering and QA managers
One of your engineers is trialling a tool you have not heard of.
You want to know what it is, what to ask them to show you, what it will cost if the team keeps it and what your security team will ask. This page answers those four, and nothing else.
- Run outcomes
- Pass rates with Wilson 95% intervals
- Fault injection recovery
- Per tool and failure mode
- Security posture
- Red-team results by category
- Judge calibration
- Per rubric
- Traced tool calls
- Trajectory and arguments
- Accessibility findings
- Automated axe-core checks
- Audit log excerpt
- Platform-wide, for the period
Your questions
Four things a manager needs before saying yes.
Not a demo. Results from one of your own agents or applications, and a pipeline that stops a bad change.
A fixed monthly price per plan. The trial ends in 14 days, and after that access pauses until someone subscribes.
Where data goes, who can read keys, whether runs can stay in your network and what evidence exists. Those have written answers.
How CertiQA helps
What to ask your engineer to show you.
A good trial ends with evidence from your own system. Here is what each item looks like in the product.
A real failure it found
Ask for one case where your agent gave a wrong answer, called a tool it should not have, or broke a policy. Tools in a test are mocked, so nothing real was touched while it was found. The run shows each tool call with its arguments, and an LLM judge scores the answer against a rubric your team wrote.
A judge is a model too. Ask how it was checked: the Release Report reports judge calibration per rubric.
Ask to see
- The failing case, as plain steps on the canvas.
- The tool calls and arguments from the run.
- The rubric the judge used.
Red-team results
What happened when the agent was attacked with prompt-injection and jailbreak scenarios from the OWASP LLM Top 10: 20+ built-in scenarios across seven categories, plus attacks generated from the agent's own tools. Ask what needs fixing before release.
Red-teaming is on Team Pro Plus and Enterprise Ultra. If the team subscribes to a lower plan, it is not kept after the trial.
Ask to see
- Results by category.
- The attacks that got through.
A CI gate that blocks a change
A failing agent test should stop a merge. Ask your engineer to make one fail on purpose and show the pipeline stopping. Then ask what the team does today for the same check and who does it.
If the trial does not produce at least the first two items, it probably is not a fit yet.
For your own review
A Release Report collects measured results into one export. When a run is signed off in the app, CertiQA generates one for that run, with the approver and decision recorded. A project-wide export for a date range is available through the API. It is evidence, not a certification.
Cost
Four plans, and a trial that ends.
The trial runs 14 days with Team Pro Plus features and 500 credits. When it ends, access pauses and CI runs stop until a plan is active.
| Plan | Per month | Per year (20% off) | Best for |
|---|---|---|---|
| Starter Solo | $299 | $2,870 | One engineer testing agents. 1 seat. |
| Team Pro | $749 | $7,190 | A team gating releases in CI. 10 seats. No red-teaming. |
| Team Pro Plus | $899 | $8,630 | Teams shipping agents. Includes red-teaming and multi-provider BYOK. 25 seats. |
| Enterprise Ultra | Custom | Private VPC runners, SAML SSO, unlimited seats. | |
Red-teaming and bring-your-own-key are in the trial but only on Team Pro Plus and Enterprise Ultra afterwards. See full plan details.
Security and procurement
What your security team will ask, and where the answers are.
Where does data go?
Secrets are encrypted at rest with AES-256-GCM. Customer data and traces are not used to train public foundation models. Subprocessors are listed on the security page.
Can runs stay in our network?
On Enterprise Ultra, runners deploy inside your network and connect outbound over HTTPS only.
Do you hold SOC 2 or ISO 27001?
No, not currently. We publish a description of our controls and a subprocessor list. Our Data Processing Addendum is published, and a workspace administrator accepts it in the product before any test can run.
How do we start a review?
Read the Data Processing Addendum, or email [email protected] to send a questionnaire. Read the security overview.
FAQ
Questions managers ask
What is CertiQA, in one paragraph?
CertiQA is a testing platform for AI agents and web applications. For agents it replaces the real tools an agent can call with controlled mocks, scores answers with an LLM judge, attacks the agent with prompt-injection and jailbreak tests based on the OWASP LLM Top 10, and can fail a CI build when something regresses. It also runs web UI and API tests, with self-healing locators for the UI.
How agent testing worksHow much does CertiQA cost?
Starter Solo is $299 a month, Team Pro is $749, Team Pro Plus is $899, and Enterprise Ultra is priced per contract. Annual billing is 20% lower. There is no permanent free plan. Red-teaming and multi-provider bring-your-own-key are on Team Pro Plus and Enterprise Ultra only.
See plans and creditsWhat does the 14-day trial include, and does it need a credit card?
The trial includes Team Pro Plus features and 500 credits. It needs no credit card and no sales call. Your engineer can invite teammates during the trial so more than one person sees the results.
What happens if we do not subscribe after the trial?
Access pauses at the end of the trial and CI runs stop, so any pipeline that calls CertiQA fails until a plan is active. If your team needs more time to evaluate, contact us.
Will a test ever touch our production systems?
Not unless someone sets it up that way. A tool with no mock is never sent anywhere. To reach a real system, a test must include an explicit step that routes that one tool to an address you choose, which is normally a staging system.
How tool mocking worksWhat will our security team ask, and what can we give them?
Typical questions are where data goes, who can see keys, whether runs can stay in your network and what evidence exists. The security page answers each and says plainly that CertiQA does not currently hold SOC 2 or ISO 27001. Our Data Processing Addendum is published, and a workspace administrator accepts it in the product before any test can run.
Security overviewWhat is a Release Report, and is it a certification?
It is a dated export of what CertiQA measured for a project for a signed-off run or a date range, with a SHA-256 hash so a reviewer can check it was not edited. It is evidence for a reviewer, not a certification. Its audit-log section covers the whole platform, not only the project.
We already use Playwright, Postman and an eval tool. Why change anything?
You may not need to. CertiQA runs Playwright underneath for UI tests and exports them as TypeScript, and it tests APIs too. The case for it is one place where agent, API and UI checks run together, with one CI gate and one report. The comparison page sets out where other tools differ.
Compare with other toolsSee also: example scenarios and all features.
Ask your engineer for two things this week.
A failure found in your own agent, and a pipeline that stops a bad change. If you would rather talk to us first, write to us.