Skip to content

MCP Deep Dive

MCP (Model Context Protocol) lets AI tools connect to external services such as databases, APIs, browsers, and documentation systems.

It is the connectivity layer, not the judgment layer. MCP tells the agent what it can reach. It does not tell the agent how to use those capabilities well.

You probably don’t need MCP if:

  • You’re just starting with AI coding
  • Your workflow is code editing + tests
  • Stock tool configuration works fine

You might need MCP for:

  • Browser automation and testing
  • Database queries and schema inspection
  • Real-time documentation lookup
  • Custom integrations

If the question is “how do I let the agent use GitHub, a browser, or a docs service?” that is probably an MCP question.

If the question is “how should the agent perform code review or release prep?” that is probably a skills question.

If you do need MCP, start with these:

MCP ServerUse CaseInstall
Context7Documentation lookupnpx -y @context7/mcp
FilesystemSandboxed file accessnpx -y @anthropic/mcp-filesystem
GitHubRepo management, PRsnpx -y @anthropic/mcp-github
PlaywrightBrowser automationnpx -y @anthropic/mcp-playwright

These are examples of MCP’s core value: one standard protocol, many tools, and less custom integration work per agent client.

The practical pattern is simple:

  1. give the agent access to the system it needs
  2. keep the tool surface narrow
  3. pair the MCP server with a workflow or skill that says how to use it

For example:

  • use GitHub MCP to inspect issues or open PRs
  • use Playwright MCP to test a UI flow
  • use Context7 MCP to fetch current docs

The connection alone is not enough. The workflow around it still matters.

MCP tools can consume massive context:

MCP ToolTypical Cost
Playwright screenshot15,000+ tokens
Full DOM snapshot10,000-50,000 tokens
Database query resultVariable

A single screenshot can exhaust hours of token allocation.

Mitigations:

  • Use targeted selectors, not full page captures
  • Limit result set sizes
  • Monitor token usage

MCP servers run with your permissions. Use caution:

Terminal window
# Sandboxed execution for untrusted servers
docker run --rm -it mcp/playwright

Use MCP when you need external access.

Use skills when you need reusable judgment.

Most serious setups use both: MCP to reach systems, skills to keep behavior consistent.