FirstQA MCP server

Run Ovi AI's pre-PR QA review inside your AI coding editor or any MCP client. Ask the agent to "review my changes with FirstQA" and it runs a functional, user-focused QA pass on your current git change (real bugs + a test recipe) before you open a PR.

It's the same engine and the same API key as the qa CLI — it just lives where you code.

Recipes identify web, api, and native surfaces. The BYO runbook includes both web and native UI scenarios. Your local coding agent executes web cases in its browser and native cases using the simulator/device tools already available in its environment. FirstQA does not require Maestro or run a simulator itself. API-only scenarios remain separately identified for API tooling or CI.

What it exposes

| Tool | What it does | |------|--------------| | firstqa_review_changes | Reviews the current git change — this branch's committed changes vs the base branch by default. Optional base, mode (branch / staged / working), and path (the repo to review). | | firstqa_run_recipe | Returns P0/P1 web and native UI scenarios as an executable local runbook. | | firstqa_submit_results | Submits scenario evidence and returns FirstQA's pass/fail/review report. | | firstqa_usage | Shows your FirstQA plan and analyses used this period. |

Which repo does it review?

MCP clients don't reliably launch the server inside the project you have open — the process working directory is often your home folder, not the repo. So FirstQA resolves the repo to review in this order, and uses the first that is a real git repository:

1. the tool's path argument,
2. the FIRSTQA_REPO_DIR env var,
3. your editor's workspace folder (MCP "roots", if the client supports it),
4. the launch directory (process.cwd()).

If none is a git repo you get a clear error listing what was tried — not a bare "Not a git repository". The fix is to pass path (e.g. "review /Users/me/project with FirstQA") or set FIRSTQA_REPO_DIR in the config. Cursor in particular often needs one of these, since it doesn't launch the server inside your workspace.

Setup

You need a FirstQA API key (fqa_live_…) — create one in FirstQA → Settings.

AI coding editor

Add the server to your editor's MCP configuration:

{
  "mcpServers": {
    "firstqa": {
      "command": "npx",
      "args": ["-y", "firstqa-mcp"],
      "env": { "FIRSTQA_API_KEY": "fqa_live_your_key" }
    }
  }
}

Any MCP client

Run npx -y firstqa-mcp as a stdio server with FIRSTQA_API_KEY in the environment.

Config

| Env var | Required | Default | |---------|----------|---------| | FIRSTQA_API_KEY | yes | — | | FIRSTQA_API_URL | no | https://www.firstqa.dev | | FIRSTQA_REPO_DIR | no | auto-detected (see Which repo does it review?) |

Usage

From your agent, just ask:

Review my changes with FirstQA before I open the PR.

The agent calls firstqa_review_changes, which collects your diff + changed files straight from git (nothing is committed or sent anywhere except FirstQA's API), runs the analysis, and returns the report — verdict, bugs, and a test recipe — right in the conversation.

Local web or mobile test run

1. Install FirstQA MCP in your coding agent and add your FirstQA API key as described above.
2. For mobile, install Xcode with an iOS Simulator or the Android SDK with an emulator. Connect whichever simulator-control tools you normally use in your coding agent; FirstQA does not require a particular tool.
3. Open the app repository and ask: "Run the FirstQA recipe against my current working changes." The run uses uncommitted changes by default. For web runs, provide the local app URL if the agent does not know it. For mobile-only runs, no URL is needed.
4. Let the agent build/launch the app, execute every listed scenario, and capture the requested screenshots and available UI snapshots/logs. It submits the evidence to FirstQA and returns the report. Unrun or weakly evidenced scenarios are reported for review, not counted as passes.

Notes

• Reads only what git reports as changed, within size caps (≤40 files, ≤100KB/file). Binary and deleted files are skipped.
• Runs count against the same monthly allowance as PR reviews and the CLI.