First-hand — our own crons & DB run through these patterns

Claude Code + MCP (2026): Give Your Agent Hands

Out of the box, Claude Code edits files and runs shell commands. MCP (Model Context Protocol) is how it graduates to operating the rest of your stack — querying your database, driving a browser, reading your issue tracker — through safe, permissioned tool servers.

What MCP actually is (30 seconds)

MCP is an open standard (started by Anthropic, now industry-wide) for connecting AI agents to tools. An MCP server is a small program exposing actions ("run SQL query", "open page and screenshot") that the agent can call — with your permission gates in between. One protocol, hundreds of servers, works across Claude Code, Cursor, Copilot and most modern agents.

Add your first server (1 command)

# example: a Postgres server, scoped to this project
claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres \
  "postgresql://readonly_user:pass@localhost:5432/mydb"

# list what's connected
claude mcp list

Restart the session and Claude Code can now answer "how many users signed up this week?" by querying your actual database — and you'll see and approve each tool call. Servers can be scoped per-project (checked into team config) or per-user.

The 5 servers worth installing first

Database (Postgres/SQLite/MySQL): the single biggest upgrade — real data answers during development instead of guessing from schema files.
Browser (Playwright): the agent can open your app, click through a flow and screenshot the bug it just fixed — closing the verify loop.
GitHub: issues, PRs and reviews without leaving the session (Claude Code also has native GitHub powers; the server adds finer-grained reads).
Filesystem (extra roots): grant access to a second repo or a docs folder outside the working directory.
Your own: an MCP server is ~50 lines with the official SDKs — wrap your internal admin API once, every agent in the company can use it.

Security rules we actually follow

Read-only credentials by default — the DB user above can't DROP anything.
Never paste secrets into prompts — servers read credentials from env/config, not from chat.
Only install servers you trust — an MCP server runs with your permissions; treat third-party ones like any dependency.
Keep permission prompts ON for writes — auto-accept reads if you like, but approve mutations.

Next steps

New to the tool itself? Start with How to Start with Claude Code. Comparing agents? Claude Code vs Cursor — both support MCP, so your servers move with you.

Disclosure & freshness: first-hand guide (2026); commands may evolve — claude mcp --help is authoritative. No affiliate links.