Skip to main content

MCP Servers

The Model Context Protocol (MCP) extends CLI agents with specialized capabilities—code research, web grounding, browser automation. While IDE-based assistants (Cursor, Windsurf) often include these features built-in, CLI agents (Claude Code, Copilot CLI, Aider) rely on MCP servers to add functionality beyond basic file operations.

These three MCP servers address the critical gaps in AI-assisted development workflows.

Code Research​

ChunkHound​

ChunkHound provides semantic code search and structured sub-agent research for large codebases.

What it does:

  • Multi-hop semantic search through code relationships
  • Hybrid semantic + symbol search (conceptual discovery, then exhaustive regex)
  • Map-reduce synthesis with architectural relationships and file:line citations

When to use it:

  • 10,000-100,000 LOC: Valuable when repeatedly connecting components across the codebase
  • 100,000+ LOC: Highly valuable as autonomous agents show incomplete findings
  • 1,000,000+ LOC: Essential—only approach with progressive aggregation at extreme scale

Key trade-off: Higher token cost (1-2x) vs autonomous search, but maintains first-iteration accuracy through context isolation.

Installation:

uv tool install chunkhound

Requires Python 3.10+ and the uv package manager. See ChunkHound on GitHub for API key configuration and setup details.

Learn more: Lesson 5: Grounding covers ChunkHound's architecture, pipeline design, and scale guidance in detail.

Web Research​

ArguSeek​

ArguSeek is a web research sub-agent with isolated context and semantic state management.

What it does:

  • Google Search API (quality vs Bing/proprietary alternatives)
  • Query decomposition—3 concurrent variations per query (docs + community + security advisories)
  • Semantic subtraction—follow-up queries skip covered content and advance research
  • Bias detection—flags vendor marketing, triggers counter-research

When to use it:

  • Researching best practices and competing approaches
  • Finding security advisories and CVEs
  • Learning new technologies with current (post-training) information
  • Multi-source research (processes 12-30 sources per call, scales to 100+ sources per task)

Key advantage: Maintains state across queries—builds on previous research instead of re-explaining basics, keeping your orchestrator context clean.

Installation:

brew install arguseek

Requires Go 1.23+ and Google API credentials. See ArguSeek on GitHub for detailed setup instructions and configuration options.

Learn more: Lesson 5: Grounding explains ArguSeek's architecture, semantic subtraction, and research patterns.

Browser Automation​

Two major options for browser automation—both provide comprehensive tooling, differ in maturity and optimization approach.

Playwright MCP​

Playwright MCP is the official browser automation server from Microsoft, built on the Playwright testing framework. Most popular MCP server on GitHub for browser automation.

What it does:

  • Accessibility tree approach (not screenshots)—LLM-friendly structured data from the DOM
  • Full browser automation via Playwright—navigate, click, type, extract data
  • Automated testing and exploration—generate tests, reproduce bugs, validate UX from natural language
  • Self-verifying workflows—agents modify code, launch browser, interact with UI, confirm expected behavior

When to use it:

  • Mature ecosystem preference—established Playwright foundation with broad community support
  • Testing-focused workflows—leverages Playwright's end-to-end testing patterns
  • Accessibility-first automation—semantic DOM structure over visual parsing

Key advantage: High popularity and mature testing ecosystem. Accessibility tree provides clean, structured text that LLMs interpret reliably without visual processing overhead.

Installation:

npx @playwright/mcp@latest

Requires Node.js 18+. See Playwright MCP on GitHub for MCP client configuration.

Chrome DevTools MCP​

Chrome DevTools MCP is the official browser automation server from the Google Chrome team, purpose-built for MCP workflows with context optimization.

What it does: (26+ professional tools)

  • Performance analysis—run traces, extract LCP, blocking time, actionable metrics
  • Advanced debugging—analyze network requests (CORS, failed loads), inspect console logs, take DOM snapshots
  • Reliable automation—simulate user interactions (click, type, navigate) via Puppeteer
  • Emulation—CPU throttling, network speed, viewport size for testing under constraints

When to use it:

  • Performance-focused workflows—deep Chrome DevTools integration for profiling and optimization
  • Context-optimized preference—newer tool designed specifically for MCP agent use cases
  • Chrome-specific features—leverage proprietary DevTools Protocol capabilities

Key capability: Closes the "write code → run → verify" loop—agents test their changes in the browser and iterate based on actual behavior.

Installation:

npx chrome-devtools-mcp@latest

See Chrome DevTools MCP on GitHub for MCP client configuration.

Choosing Between Them​

Playwright MCP: More popular with broader GitHub community, mature testing ecosystem, established Playwright foundation. Best for standard testing workflows and accessibility-first automation.

Chrome DevTools MCP: Newer and purpose-built for MCP, context-optimized by the Chrome team, performance analysis focus. Best for Chrome-specific debugging and profiling workflows.

Both provide comprehensive browser automation with similar scope (~26 tools). The choice depends on ecosystem preference and whether you prioritize maturity (Playwright) or MCP-specific optimization (CDP).

Run Browser Automation in Sub-Agents

Browser automation generates high token volumes—DOM snapshots (5,000-15,000 tokens), screenshots (3,000-8,000 tokens), network traces (2,000-10,000 tokens). Multiple operations quickly fill your context window.

Best practice: Delegate browser tasks to sub-agents. The sub-agent processes DOM data and screenshots in its isolated context, then returns a concise synthesis: "Button at selector .submit-btn clicked, form submitted successfully, redirected to /dashboard" (50 tokens instead of 15,000-token DOM dump).

See Lesson 5: Sub-Agents for Context Isolation for architecture details.

Related Course Content: