Generated suites arrive in pull requests with a lot of code. Reviewers skim, and the weak assertions get through.
For teams using Cursor, Claude Code or Claude Desktop
Your assistant can write tests faster than anyone can review them.
The first draft of a test suite is no longer the slow part. The slow part is a reviewer reading hundreds of lines of generated code to decide whether the tests check anything that matters. This is how CertiQA moves that review onto plain steps.
| Client Application | Authorized By | |
|---|---|---|
| Cursor IDE | maya@acme.test | Disconnect |
| Claude (Anthropic) | maya@acme.test | Disconnect |
Example data. Mirrors the connected clients table in Settings, API Keys. It also shows scopes and dates.
What this looks like day to day
Four worries that come with AI-written tests.
"The page loads" passes forever. You want to see the expected order total, the specific message, the actual failure path.
The people who know whether the checkout rules are right often do not read TypeScript. They are left out of the review.
Giving a tool your credentials is a decision. You want to know what it can change and how to take that access back.
How CertiQA helps
Each worry has a specific answer.
The assistant writes into CertiQA, not into your repository
Connected over MCP, the assistant calls CertiQA's tools to create projects, suites, scenarios and cases. They appear in the project on the visual canvas as a list of steps: do this, expect that. For an AI agent under test, a case shows its conversation turns, mocked tool responses and expected tool calls.
A reviewer reads, edits and approves the steps, and runs the suite. A web UI case can still be exported as Playwright TypeScript from the app if you want the code.
An assistant's output is a draft. It can mark each step as AI-written with a confidence, and the lowest-confidence steps are the ones to read first, but the review is still yours.
Using the certiqa tools, list my
projects. In the Storefront project,
create a suite called "Checkout - guest"
with cases for: successful card payment,
declined card, expired promo code, and
an out-of-stock item at the last step.
Do not run anything yet.Connect once, without copying a key
Open Settings, API Keys, and find the Connected AI Clients & Remote MCP card. It shows your remote MCP server URL, with tabs for Claude Desktop, Cursor and a terminal installer. For Claude Desktop and Cursor, one command writes the configuration:
npx -y @certiqa/mcp-server install --client cursor --url <your MCP URL>
Restart the client and sign in to CertiQA when it prompts. That is an OAuth sign-in, so there is no API key to copy. Claude Code adds the same URL with its own claude mcp add command, and a local server with a scoped API token is available if you prefer it.
Connecting
- Claude Desktop and Cursor: one installer command.
- Claude Code: its own add command.
- Sign in with OAuth, then restart the client.
Read the steps, not the code
Each step is a plain action with an expected result, so a QA lead, product owner or domain expert can judge it. A short checklist helps: does each case protect a behavior you care about, are assertions specific, are the data and locators right, and for agent cases do the mocked responses cover the failure paths.
Edit the steps directly on the canvas, or ask the assistant to revise them.
What the 36 tools cover
- Projects, suites and scenarios.
- Creating cases and agent cases, and adding them to suites.
- Validating step syntax before saving.
- CertiBlocks: list, read, create, add steps.
- Triggering runs, agent suite runs and schedules.
- Reading run reports, failures, and repairing case steps.
Know what it can reach, and switch it off
A connected assistant acts as you, limited to the scopes you grant at sign-in. Reading needs a read scope, starting runs needs a run scope and other changes need a write scope. An OAuth connection cannot reach billing, members, API tokens, vault secrets or integration settings, and cannot delete projects. There is no tool for deleting a project or deleting a case (deleting is done in the app).
In Settings, the connected clients table has a Disconnect button that revokes a client's access immediately.
The assistant can still create cases, create schedules and start runs, and runs use credits. Ask it to hold off running anything until you have read its draft.
- Every plan and the 14-day trial
- Author with your own assistant
- Team Pro Plus, Enterprise Ultra
- Full MCP suite for headless CI
- Team Pro Plus, Enterprise Ultra
- Cloud triage agents on CI failures
FAQ
Questions about authoring tests with an assistant
Can Cursor or Claude Code write tests in CertiQA?
Yes. Connect Cursor, Claude Code or Claude Desktop to CertiQA's MCP server and the assistant can create projects, suites, scenarios and test cases, including AI agent cases with mocked tool responses and expected tool calls. The cases appear in the project as plain steps on the visual canvas for your team to review.
How do I connect an AI assistant to CertiQA?
Open Settings, then API Keys, and find the Connected AI Clients & Remote MCP card. It shows your remote MCP server URL. For Claude Desktop and Cursor, one command writes the configuration for you. Restart the client and sign in to CertiQA when it prompts. Claude Code uses its own claude mcp add command with the same URL. The setup guide has the exact commands for each client.
Do I need to copy an API key?
Not for the recommended setup. The remote server uses OAuth, so you sign in as yourself and the assistant never holds a long-lived key. If you prefer a shared or scripted setup, you can instead run the local MCP server with a scoped API token from Settings. Keep that token in a secret manager or environment variable, not in a file you commit.
What can the assistant do through MCP?
The MCP server has 36 tools. They cover listing and creating projects, suites and scenarios; creating test cases and agent test cases and adding them to suites; listing, reading and creating CertiBlocks; validating case steps before saving; triggering runs and agent suite runs; building scripted conversation checks and adding them to an existing agent case; comparing a normal run with a constrained or keyboard-only run; creating and deleting schedules; and reading run status, reports, failures and artifacts for triage. They include reading a case and replacing its steps, so an agent can repair a case after a failed run; deleting a case is done in the app.
Why not just let the assistant write Playwright code?
Generated code has to be read line by line before you can trust it, and that review is the part that gets skipped. Here the assistant writes plain steps and assertions into CertiQA, so a QA lead, product owner or domain expert can read, correct and approve them without reading code. You can still export a web UI case to Playwright TypeScript from the app afterwards.
Web UI automationHow much access does a connected assistant get?
It acts as you, limited to the scopes you grant when you sign in. Reading needs a read scope, starting runs needs a run scope, and other changes need a write scope. An OAuth connection cannot reach billing, members, API tokens, vault secrets or integration settings, and cannot delete projects. In Settings the same card lists connected clients with a Disconnect button that revokes access immediately. Ask the assistant not to run anything until you have reviewed its draft.
Security overviewDoes this cost extra, and which plans include it?
Authoring tests from Cursor, Claude Code or Claude Desktop is available on every plan and during the 14-day trial, using your own Cursor or Claude subscription for the drafting. Running the tests it creates uses credits like any other run. The full MCP suite for headless CI orchestration, and cloud triage agents on CI failures, is part of Team Pro Plus and Enterprise Ultra.
See plansCan the assistant reuse my existing login or setup blocks?
It can list your CertiBlocks, read one, create a new one and add steps to it. The tool that creates cases takes ordinary steps and has no block-insert action, so to use a block inside a case, insert it in the app.
Reusable test blocksTry it on one small feature.
Connect your assistant, ask for a draft suite and tell it not to run anything yet. Then read the steps it made.