cypress-mcp

MCP server that shows an AI assistant what happened in your last cypress run: which commands each test ran, what failed, the error, a screenshot, and the page as it was at the moment of the failure.

It is made for headless runs, on your machine and in CI. That is the place the official tools do not reach, see How it compares.

npm install -D @miroslavmyrha/cypress-mcp

Example

A spec with three tests, the last of which fails:

it('shows a counter of finished items', () => {
  cy.get('.todo-list li').first().find('input[type=checkbox]').check()
  cy.get('[data-test=finished-count]').should('contain', '1')
})

After cypress run, the assistant calls get_last_run and gets the commands that ran, the error and the path of a DOM snapshot (output shortened):

{
  "title": "Todo list > shows a counter of finished items",
  "state": "failed",
  "domSnapshotPath": "snapshots/cypress-e2e-todo-cy-js/todo-list-shows-a-counter-of-finished-items-d963bb8d.html",
  "commands": [
    { "name": "visit", "message": "https://example.cypress.io/todo" },
    { "name": "get", "message": ".todo-list li" },
    { "name": "first", "message": "" },
    { "name": "find", "message": "input[type=checkbox]" },
    { "name": "check", "message": "" },
    { "name": "get", "message": "[data-test=finished-count]" }
  ],
  "error": "AssertionError: Timed out retrying after 2000ms: Expected to find element: `[data-test=finished-count]`, but never found it."
}

It can then look into the page as it was when the test failed. query_dom with the selector .todo-count:

{
  "selector": ".todo-count",
  "totalMatches": 1,
  "results": [
    {
      "breadcrumb": "body > section.todoapp-wrapper > div.todoapp > footer.footer > span.todo-count",
      "outerHTML": "<span class=\"todo-count\"><strong>1</strong> item left</span>"
    }
  ]
}

So the counter exists, under another selector, and the assistant can fix the test without anyone pasting logs into the chat.

How it works

cypress run → tests run
  → support/index.ts captures Cypress command log per test
  → cy.task('mcpSaveTestLog') transfers logs to node side
  → plugin/index.ts stores logs in memory
  → after:spec → merges logs + results → writes .cypress-mcp/last-run.json

Claude → MCP server → get_last_run → reads last-run.json

How it compares

Cypress has two tools of its own for AI assistants. They cover other situations than this one.

| | this package | cypress tap | Cypress Cloud MCP | |---|---|---|---| | Reads | the last local cypress run | a live cypress open session | runs recorded to Cypress Cloud | | Headless runs and CI | yes | no | only what was recorded | | Command log and DOM of a failed test | yes | yes | no, links to Test Replay | | Needs | Cypress 12 or newer, a plugin and a support import | Cypress 15.21 or newer, a Chromium browser | a Cypress Cloud account | | Interface | MCP (stdio or HTTP) | command line | MCP |

Use cypress tap while you write a test with the Cypress app open. Use this package when the tests run headless and the assistant should see why one failed.

The results are written in the after:spec event, which Cypress fires in run mode. In cypress open that event needs experimentalInteractiveRunEvents.

Before anything reaches the assistant, values that look like secrets are redacted, and content that comes from the tested application is marked as untrusted data.

Setup

1. Install

npm install -D @miroslavmyrha/cypress-mcp

Note: the unscoped cypress-mcp name on npm belongs to an unrelated project. Always use the scoped @miroslavmyrha/cypress-mcp.

2. Register the plugin in cypress.config.ts

import { defineConfig } from 'cypress'
import { cypressMcpPlugin } from '@miroslavmyrha/cypress-mcp/plugin'

export default defineConfig({
  e2e: {
    setupNodeEvents(on, config) {
      cypressMcpPlugin(on, config)
    },
  },
})

3. Import support in cypress/support/e2e.ts

import '@miroslavmyrha/cypress-mcp/support'

4a. Configure Claude Desktop / Claude Code (stdio transport)

In your MCP settings (~/.config/claude/claude_desktop_config.json or claude mcp add):

{
  "mcpServers": {
    "cypress": {
      "command": "npx",
      "args": ["-y", "@miroslavmyrha/cypress-mcp", "--project", "/absolute/path/to/your/project"]
    }
  }
}

Or with claude mcp add:

claude mcp add cypress -- npx -y @miroslavmyrha/cypress-mcp --project /absolute/path/to/your/project

4b. HTTP transport (Ollama, custom bridges)

# Start MCP server in HTTP mode (random token printed to stderr)
npx -y @miroslavmyrha/cypress-mcp --project /path/to/project --transport http --port 3333

# Or with a fixed token for automated deployments
MCP_HTTP_TOKEN=$(openssl rand -hex 32) npx -y @miroslavmyrha/cypress-mcp --project /path/to/project --transport http --port 3333

# Server listens on:
#   POST http://localhost:3333/mcp  → tool calls (MCP Streamable HTTP)

MCP Tools

| Tool | Description | |------|-------------| | list_specs | List all Cypress spec files. Optional pattern glob (default: **/*.cy.{ts,js,tsx,jsx}) | | read_spec | Read a spec file's content by relative path | | get_config | Read the Cypress config file (cypress.config.ts/js/mjs/cjs), secret-looking values redacted | | get_last_run | Get results of the last test run: states, errors, command log, DOM snapshots | | get_screenshot | Get a screenshot from the last run. Returns the image itself (up to 2 MB) plus path and size metadata | | query_dom | Query the DOM snapshot of a failed test using a CSS selector | | run_spec | Run a single Cypress spec file and wait for completion. Returns exit code and summary |

Output format of .cypress-mcp/last-run.json

{
  "timestamp": "2026-02-18T22:00:00Z",
  "specs": [
    {
      "spec": "cypress/e2e/login.cy.ts",
      "stats": { "passes": 3, "failures": 1, "pending": 0, "skipped": 0, "duration": 5230 },
      "screenshots": ["/abs/path/to/Login -- should show error (failed).png"],
      "tests": [
        {
          "title": "Login > should login with valid credentials",
          "state": "passed",
          "duration": 1200,
          "error": null,
          "domSnapshotPath": null,
          "commands": [
            { "name": "visit", "message": "/login" },
            { "name": "get", "message": "#email" },
            { "name": "type", "message": "[email protected]" }
          ],
          "consoleErrors": [],
          "networkErrors": []
        },
        {
          "title": "Login > should show error on invalid password",
          "state": "failed",
          "duration": 800,
          "error": "Timed out retrying: Expected to find element: .error-message",
          "domSnapshotPath": "snapshots/cypress-e2e-login-cy-ts/login--should-show-error-a1b2c3.html",
          "commands": [...],
          "consoleErrors": ["Uncaught TypeError: Cannot read property 'x' of null"],
          "networkErrors": [{ "method": "POST", "url": "/api/login", "status": 500 }]
        }
      ]
    },
    {
      "spec": "cypress/e2e/dashboard.cy.ts",
      "stats": { ... },
      "screenshots": [],
      "tests": [...]
    }
  ]
}

All specs from a run are captured. In cypress run the accumulator resets on before:run. In cypress open results accumulate for the lifetime of the Cypress process. Failed tests include a domSnapshotPath pointing to an HTML file (up to 100 KB of document.body.outerHTML) stored under .cypress-mcp/snapshots/.

CLI options

Options:
  --project <path>     Cypress project root (default: current directory)
  --transport <type>   stdio or http (default: stdio)
  --port <number>      HTTP port, only for --transport http (default: 3333)
  -V, --version        Show version
  -h, --help           Show help

Add .cypress-mcp to .gitignore

.cypress-mcp/

Environment variables

| Variable | Description | |----------|-------------| | MCP_HTTP_TOKEN | Set a fixed bearer token for HTTP transport (min 32 chars). If not set, a random token is generated on each start and printed to stderr. |

Development

This project uses Bun as the package manager:

bun install
bun run typecheck
bun run test
bun run build

Requirements

• Node.js 18+
• Cypress 12+ (developed and tested with Cypress 13)

Author

Miroslav Myrha

License

MIT