Local-first · no cloud account · MCP-native

Record once.
Replay as a ghost.

The local-first QA agent for SaaS teams. Record browser flows once, replay them in CI, monitor staging, and let AI assistants run your test suite via MCP — all on your machine.

Get started View on GitHub

npm install -g ghostrun-cli

Why GhostRun

Nobody else combines these four things

Playwright MCP and Stagehand automate a browser. Reflect, Momentic, and Mabl want an account and a hosted dashboard. GhostRun is the one that's actually local, actually SaaS-specific, and actually speaks to your AI agent.

Genuinely local-first

Every app repo gets its own .ghostrun/ workspace. Secrets stay in env vars and CI secrets — never in the npm package, never in a third party's cloud.

Built for SaaS staging

OTP bypass for phone/WhatsApp auth. db:assert against staging Postgres for multi-tenant RLS checks. Webhook signature verification for Razorpay, Meta, and custom HMAC.

MCP-native, not bolted on

A real MCP server with tools for running suites, inspecting failures, and reviewing repair proposals — not just raw click/type primitives.

Self-healing you can trust

Selector repairs are reviewable proposals, never silent mutations. Nothing changes your flow graph in CI without a human saying yes.

How it works

One workspace. Browser, API, and load — together.

Record Click through a real flow once against staging. GhostRun captures selectors, assertions, and API calls.
Replay Run it headless, locally or in CI, against any profile — staging, preview, or production-shadow.
Watch it heal When a selector breaks, GhostRun proposes a fix. You review it — or gate CI on it never applying silently.
Hand it to an agent Connect Claude Desktop or Cursor via MCP. Ask it to run the smoke suite and summarize what broke.

~/your-saas-app

# record a flow against staging
$ ghostrun learn https://staging.yourapp.com

# run it in CI
$ export STAGING_QA_PASSWORD='...'
$ ghostrun run smoke --profile staging --ci --reporter junit

✓ checkout-guest       passed · 2.1s
✓ login-basic          passed · 0.8s
✓ signup-verify        passed · 3.4s

Key commands

A CLI, not a dashboard

Command What it does

ghostrunZero-config home menu

ghostrun learn <url>Record a flow

ghostrun explore <url>Auto-discover flows via BFS crawl + AI

ghostrun run <id> --ciRun a flow in CI mode

ghostrun suite:run smokeRun a named suite

ghostrun repair listReview self-healing proposals

ghostrun auditScan for secret leaks before committing

ghostrun doctorHealth check

Agent-native

Your AI assistant can run the whole thing

Connect Claude Desktop or Cursor and ask it to operate GhostRun directly — run flows, inspect failures, review repairs — through a real MCP server, not a screenshot loop.

claude · mcp · ghostrun

› Run the smoke suite against staging and tell me
  what broke, with a proposed fix if there's a selector issue.

✓ suite:run smoke --profile staging --ci
✓ 7 passed, 1 failed — checkout-guest
✓ repair proposal: selector drift on "#submit-btn"
  → awaiting your review, nothing applied

{
  "mcpServers": {
    "ghostrun": {
      "command": "ghostrun-mcp",
      "cwd": "/path/to/your-saas-app"
    }
  }
}

Pairs with

GhostRun drives the app. Atlas simulates the world it depends on.

Most flaky SaaS tests fail because of the providers behind the app — Razorpay retries, WhatsApp rate limits, delayed webhooks. Atlas is a stateful simulator for exactly those dependencies. Point your staging app at Atlas, then let GhostRun click through the UI and assert on what happens when a payment fails three times in a row.

Explore Atlas →

Get started

Drop it into CI in one job

.github/workflows/qa.yml

jobs:
  qa:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - run: npm ci && npx playwright install chromium
      - run: ghostrun suite:run smoke --profile staging --ci --reporter junit
        env:
          STAGING_QA_PASSWORD: ${{ secrets.STAGING_QA_PASSWORD }}

Read the docs Contribute npm package

MIT licensed. Node ≥ 20. No account, no telemetry you didn't ask for.

Record once. Replay as a ghost.