Test from Gemini CLI, the Claude desktop app and claude.ai
Install Testsy as a Gemini CLI extension or a Claude connector, then ask in plain words for a browser or phone test and get a verdict with evidence.

Not everyone builds in an IDE#
Plenty of people who build with AI never open Cursor. They use a terminal agent such as Gemini CLI, or they work in the Claude app and ask it to plan, write and ship. They still need the same thing: a way to check that what was built works for a user, without trusting the model that built it.
Testsy's MCP server works in both. These are the setups from the MCP docs.
Gemini CLI#
gemini extensions install https://github.com/testsy-ai/testsy-gemini-extensionIt asks for your API key when you install it (create one in Testsy under Settings → API keys). The extension brings the MCP connection and the three skills: testsy-check-release, testsy-investigate-run and testsy-regression-from-bug.
The Claude desktop app and claude.ai#
Two options:
- Under Customize, add
testsy-ai/testsy-claude-pluginsas a marketplace and install Testsy from it. - Or add
https://app.testsy.dev/mcpas a custom connector under Customize → Connectors and sign in.
With the connector, there is no key to copy: you sign in to Testsy in the browser, choose the workspace and approve what the client may do. Testsy asks for one of three levels: read results, run saved tests, or also create and edit tests. Pick the least your workflow needs; you can remove the connection in Account settings → Connected apps.
For Claude Code in the terminal, see Make Claude Code test what it just built.
What to ask for#
Be specific about what a user does and what they should see:
The agent drafts a plan with plan_test and shows it to you. After you confirm, it saves the test, starts a run and hands you a watch link you can open to see the browser live. When the run ends it reads the verdict and each check's evidence. For a failed check you get what was expected and what was seen; for a passed one, the value it was decided on.

Good to know#
- The tester runs in our cloud, so the app must be reachable from the internet. A site only on your laptop cannot be tested yet.
- Runs can sign up, create records and send email. Use staging and made-up details.
- Your client's approval settings decide when it asks before a tool runs.
The free trial has 10 runs and needs no card.
This post is part of Testing an app you built with AI.
Filed underWebAI-built appsCI, schedules and agents