Automation

MCP server

Let Claude Code, Cursor and other agents write and run tests.

All documentation

The MCP server lets coding agents such as Claude Code, Cursor, VS Code with Copilot and Windsurf write tests for the features they build, run them in a real browser, and read the failures, including the failure screenshot. It is part of the Pro and Team plans.

Connect

Create an API key in Settings, API keys and MCP. The server URL is https://abackinspect.com/mcp and uses the Streamable HTTP transport with the key as a bearer token.

Claude Codebash

claude mcp add --transport http aback-inspect https://abackinspect.com/mcp --header "Authorization: Bearer aik_..."

.cursor/mcp.jsonjson

{
  "mcpServers": {
    "aback-inspect": {
      "url": "https://abackinspect.com/mcp",
      "headers": {
        "Authorization": "Bearer aik_..."
      }
    }
  }
}

.vscode/mcp.jsonjson

{
  "inputs": [
    {
      "type": "promptString",
      "id": "aback-inspect-key",
      "description": "Aback Inspect API key",
      "password": true
    }
  ],
  "servers": {
    "aback-inspect": {
      "type": "http",
      "url": "https://abackinspect.com/mcp",
      "headers": {
        "Authorization": "Bearer ${input:aback-inspect-key}"
      }
    }
  }
}

Tools

Tool	What it does
`get_test_guide`	Read this before writing tests: the definition format, every step type, locator advice, template tokens, examples and limits.
`list_projects`	Projects in the workspace with their environments (base URL, allowed hosts, verification) and test health.
`create_project` (changes data)	Create a project for a website. Only for sites the user owns or is allowed to test.
`list_tests`	Tests with their last result. Filter by project.
`get_test`	A test's current definition and metadata.
`validate_test`	Check a definition without saving it. Returns every problem with its path.
`create_test` (changes data)	Save a new test in a project. Validate first with validate_test.
`update_test` (changes data)	Change a test's name, tags, enabled state or definition. A changed definition creates a new version.
`run_test` (changes data)	Run one test in a real browser and, by default, wait for the result. Pass base_url to test a preview deployment whose host is in the environment allow list.
`run_project` (changes data)	Run every enabled test in a project (or those with a tag) and wait for the results.
`get_run`	Step results with expected and actual values, console errors, failed requests and links.
`get_run_screenshot`	Returns the failure screenshot of a run, or the screenshot of a given step, as an image.
`list_runs`	Recent runs, newest first. Filter by project, test or status (passed, failed, active, failing).
`cancel_run` (changes data)	Stop a queued or running run.

Prompts that work well

"Write an Aback Inspect test for the checkout you just built in project Shop, run it against staging, and fix the code until it passes."
"Run the smoke tag of project Shop against https://pr-142.preview.example.com and tell me what failed."
"Look at the last failed run of the sign in test, explain the cause from the step results and the screenshot, and propose a fix."

Good to know

Runs started by agents count towards your monthly runs like any other run, and are labelled MCP in the history.
The key's scopes apply: a key without the write scope cannot create or change tests.
To test a preview deployment, pass base_url; its host must be in the environment's allowed hosts, for example *.preview.example.com.
Tests are data in a validated format. An agent cannot run code in the test browser beyond the documented steps, and no tool returns secret values. A key with the write scope can still write a test that types a secret into a visible field and screenshots it, so only give agents keys for workspaces whose secrets belong to test accounts.