Lastest

@lastest/mcp-server

MCP server for Claude Code, Cursor, Windsurf, Cline, and Claude Desktop. Lets AI agents run visual regression tests, review screenshot diffs, and approve baselines on a Lastest instance through the Model Context Protocol.

Quick start

npx -y @lastest/mcp-server@latest --url https://your-lastest-instance --api-key YOUR_API_KEY

Generate an API key in the Lastest UI: Settings → Runners & API Access → Create API Key (shown only once). The server speaks MCP over stdio — wire it into any compatible client below.

Compatible clients

Any MCP-compliant client that can launch a stdio server with arguments works.

Install — Claude Code

claude mcp add lastest -- npx -y @lastest/mcp-server@latest \

--url https://your-lastest-instance \ --api-key YOUR_API_KEY

Verify with claude mcp list.

Install — Claude Desktop / Cursor / Windsurf / Cline

Add to claude_desktop_config.json, ~/.cursor/mcp.json, or your client's MCP config:

{

"mcpServers": { "lastest": { "command": "npx", "args": [ "-y", "@lastest/mcp-server@latest", "--url", "https://your-lastest-instance", "--api-key", "YOUR_API_KEY" ] } } }

Restart the client.

What an agent can do

• List repositories, tests, functional areas, builds, and active jobs
• Create AI-authored tests from a URL or natural-language prompt
• Update or delete tests and functional areas
• Self-configure test runtime: playwrightOverrides, diffOverrides, stabilizationOverrides, viewport, setup wiring
• Manage shared auth: storage states (Playwright storageState() blobs) and reusable setup scripts
• Trigger test runs (optionally forceVideoRecording) and read run results
• Inspect visual diffs and approve / reject baselines individually or in bulk
• Heal failing tests with AI based on the latest run
• Publish, list, and revoke public-share links for builds and tests
• Read and update repo-level Playwright settings (browser, viewport, timeouts, parallelism, stabilization)
• Drive the Verify phase: Change Map, step comparisons, per-layer approvals
• Pull QA summaries, build review reports, and coverage stats

Tools exposed

The server registers 20 MCP tools (all prefixed lastest_). Every tool returns a structured { status, summary, actionRequired?, details } payload.

Pure CRUD operations are consolidated into resource tools that take an action (or scope) discriminator; workflow verbs stay as standalone tools.

Resource tools (action-dispatched)

Workflow verbs (standalone)

Self-configuring tests

lastest_test (action: "update") accepts a full override surface so an agent can shape a test without touching the UI:

• Setup wiring — setupTestId (use another test as setup, takes precedence) or setupScriptId, plus setupOverrides / teardownOverrides blocks to inject/skip default steps (test | script | storage_state).
• Runtime overrides — playwrightOverrides (browser, navigation/action/selector timeouts, error modes, baseUrl, cursor speed), viewportOverride, diffOverrides, stabilizationOverrides.
• Lifecycle — quarantined, executionMode (procedural | agent).

Pass null to any override block to clear it. The API validates each referenced id is in the same repo before persisting.

Migration — tool consolidation (v0.3.x)

The tool surface was consolidated from ~56 discrete tools to ~21 by folding pure CRUD into the resource tools above. This is a breaking change for scripts that called the old per-operation tool names. The workflow verbs and the QuickStart chain (lastest_quickstart → lastest_quickstart_status → lastest_publish_share) kept their names and signatures. Map any removed names as follows:

If you need the old names temporarily, pin a 0.2.x release of this package while you migrate.

CLI

lastest-mcp --url <url> --api-key <key>

The process communicates with the host client over stdio.

Requirements: Node.js 18+ and a reachable Lastest instance.

Remote (Streamable HTTP)

Most people want the endpoint your Lastest instance already serves:

https://your-lastest-instance/api/mcp

It speaks Streamable HTTP and accepts either an API key or an OAuth 2.1 connection, so an agent platform that only