A human-in-the-loop QA runner: an AI works through the test sheet on real devices, and a person signs off each case.
When something ships broken, "the AI checked it" is not an answer a client or an auditor will accept. The git-qa MCP server and its desktop app are built on that premise. The agent does the tedious part, stepping through a regression sheet on a real device, and a person watches the live screen and records the verdict for each case with a single key.
The point of the tool is a distinction most test reports blur. A plain "pass" can mean a person looked, or that nobody did. git-qa keeps those apart:
| Verdict | Meaning |
|---|---|
VERIFIED |
A person put their handle on it |
AUTO_PASS |
The AI passed it and nobody looked |
FAIL |
It failed |
BLOCKED |
A precondition was not met |
SKIP |
Not run |
Pressing the key is a signature, not a lock. You can mark a case verified without looking; your name simply goes into the evidence either way.
Android phones and emulators over adb; web pages in Chrome, Edge, Firefox or Safari at a fixed viewport; desktop apps on macOS (Accessibility plus Vision OCR) and Windows (UI Automation); and iPhone or iPad over USB, with taps and typing going through a WebDriverAgent build you sign and install yourself. Android Chrome works through port forwarding, though text inside input fields is not read back.
The git-qa MCP server lets Claude Code or another agent screenshot the target, read what is on screen, list the names of things that can be pressed, and tap one by name, on any of those platforms. That is useful when an agent has just built a feature and should check it on a real device. A run can also go unattended overnight with --no-ui, leaving the failures for you to review in the morning. The final pass is still a human decision.
Download the desktop app for macOS or Windows (x64 or ARM64), or run it from a clone with pnpm. Node 22 or newer is required and must be installed system-wide, because an app launched from an icon cannot see a Node managed only by fnm or nvm in your shell. You also need adb for Android and a browser for web targets; ffmpeg and cwebp are optional and only compress the evidence. pnpm doctor reports what your machine can actually do. The MCP server runs from the clone; there is no npm package.
Each run writes a folder with run.json holding verdicts and who placed them, plus a screenshot per case and video only when requested. Paths are relative, so the folder can be committed to Git.
For browser automation without the sign-off model, see Playwright or Chrome DevTools, and browse more developer tools.
Read, search and edit files inside allowed directories with the Model Context Protocol reference Filesystem server.
Retrieve version-aware library documentation and code examples with Context7 MCP, using resolve-library-id and query-docs.
MCP server implementation for using Claude API with Claude Desktop, providing advanced API integration and conversation management.
Query threat intelligence reports, CVEs, IOCs, STIX bundles, and weekly briefings — all through natural language, powered by the [Model Context
Formula One MCP server using FastF1
This MCP server provides integration with Gerrit code review system, allowing AI assistants to review code changes and their details through a simple