MCP — Model Context Protocol

MCP (Model Context Protocol) is an open protocol that lets AI models use external tools through a standardized interface. Instead of an AI model being limited to generating text, MCP lets it call functions — read files, click buttons in a browser, run queries — via a consistent request/response contract.

In the Leadmetrics context, MCP is used in two ways:

Development tooling — Claude Code’s Playwright MCP server is used to automate browser interactions during development, QA, and E2E testing.
Platform integration (planned) — Leadmetrics can expose its own MCP server so external AI systems (like a client’s Claude-based assistant) can query campaign status, trigger content generation, or read insights by calling structured tools.

Playwright MCP (Development Use)

The Claude Code desktop app ships with a built-in MCP Playwright server. It lets Claude control a browser directly — navigate, click, fill forms, take screenshots, and read the DOM — without needing a Selenium wrapper or separate test runner.

How to use it

The Playwright MCP tools are available when using Claude Code with the MCP Playwright integration enabled. Tools include:

Tool	What it does
`browser_navigate`	Navigate to a URL
`browser_snapshot`	Capture the current page’s accessible element tree
`browser_take_screenshot`	Take a screenshot
`browser_click`	Click an element by ref
`browser_fill_form`	Fill a form with field values
`browser_type`	Type text into a focused element
`browser_evaluate`	Run arbitrary JavaScript in the page context
`browser_wait_for`	Wait for a selector, URL, or time
`browser_network_requests`	Inspect outgoing network requests

Windows-specific guidance

On Windows, always use http://127.0.0.1:PORT rather than http://localhost:PORT. The browser sends Origin: http://127.0.0.1:3000 which differs from localhost — the Leadmetrics API CORS config accepts both.

Best practices

Take a fresh browser_snapshot after every navigation — element refs go stale across page loads.
Use browser_evaluate with document.querySelectorAll + text matching for robust element targeting rather than relying on snapshot refs.
For React-controlled inputs, use the nativeInputValueSetter pattern via browser_evaluate — direct .value = assignment bypasses React’s synthetic events.
Keep open tabs to ≤ 2 to reduce crash risk.
Avoid window.scrollTo via browser_evaluate — it can crash the browser session. Navigate directly to the target URL instead.

See Known Issues for a full list of Windows-specific bugs and workarounds.

Leadmetrics MCP Server (Planned)

A future apps/servers/mcp service will expose Leadmetrics capabilities as MCP tools, allowing external AI assistants to interact with the platform programmatically.

Proposed tools:

Tool	Description
`get_tenant_overview`	Returns campaign summary, deliverable counts, pending approvals
`list_activities`	Lists activities with status filter
`get_activity`	Returns full detail for one activity
`approve_content`	Approves a content activity (requires auth)
`generate_blog`	Triggers the blog writer agent for a given content brief
`get_insights`	Returns latest channel insights for a tenant
`search_knowledge_base`	Searches the tenant’s RAG knowledge base

Authentication will use the same JWT system as the existing API. Each MCP client authenticates with a tenant-scoped API key.

Known MCP Issues — Windows Playwright MCP bugs and workarounds
Agent Chat — The in-app AI chat that uses a similar tool-calling pattern
Adapters — How Leadmetrics agents call external LLM runtimes