Without MCP, Claude can think. With MCP, Claude can do — pull data from Linear, open pages in your browser, query your database, comment on GitHub PRs. The difference between Claude-as-assistant and Claude-as-agent.
Most Claude Code users skip MCP for months. It looks like config-file fiddly-bits. But once you have two or three servers wired up, the productivity unlock is real — Claude starts doing the parts of your job that used to require context-switching across five apps.
This guide explains what MCP actually is, how to install a server in 2 minutes, and the five servers worth installing on day one.
What MCP is (and isn't)
MCP — Model Context Protocol — is a small open protocol. A "MCP server" is just a tiny program (usually run via npx) that exposes a set of capabilities Claude can call. It runs locally on your machine. Claude Code talks to it over standard input/output.
What MCP servers expose:
- Tools — functions Claude can call (e.g., "get-customer-by-email")
- Resources — files or data Claude can read (e.g., the contents of a Notion page)
- Prompts — templated prompts the user can invoke
The user-facing result: Claude can do things in your tools. Not just read about them.
How to install an MCP server
Edit ~/.claude/mcp.json (create it if it doesn't exist):
// Add servers here. Restart Claude Code to pick them up. { "servers": { "github": { "command": "npx", "args": ["-y", "@mcp/server-github"], "env": { "GITHUB_TOKEN": "ghp_…" } } } }
Restart Claude Code. Now ask Claude something that requires GitHub access — "What PRs are open in this repo?" — and it'll call the GitHub MCP server, fetch the data, and answer.
Two-minute setup. The hard part is picking the right servers.
The five worth installing first
1. Filesystem MCP
What it does: lets Claude read and write files outside the current working directory. By default Claude Code is sandboxed to the repo you ran it in; Filesystem MCP opens up specific directories you allow.
Why it's essential: cross-repo work, accessing notes in a separate folder, working with downloaded data files. Without it you're re-explaining context every session.
Install:
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/Documents", "/Users/you/notes"]
}2. GitHub MCP
What it does: read/write GitHub via the API. PRs, issues, releases, comments, repository contents.
Why it's essential: Claude can review PRs across your org, file issues from your description, find code across repos, write changelogs from merged PRs.
Generate a personal access token at github.com/settings/tokens. Scope it to what you need (read at minimum; write if you want Claude to comment).
3. Browser / Chrome MCP
What it does: drives a real Chrome browser. Open URLs, click things, fill forms, read pages.
Why it's essential: any task that requires logging into a service Claude can't access via API. Internal admin panels. Personalized dashboards. SaaS tools without good APIs.
Anthropic publishes one (claude-in-chrome). There are also community options. Pick one and don't look back — once you can automate browser tasks via Claude, you stop doing them yourself.
4. Linear or Jira MCP (whichever you use)
What it does: read and manage your issue tracker. Create issues, update status, link to PRs, list sprint items.
Why it's essential: turns "tell me what's in sprint 12" from a context-switch into a one-line question. Even better, turns "file an issue about X" from a 5-minute task into a 5-second one.
For Linear, install the official MCP server. For Jira, the community Atlassian MCP server works well.
5. Database MCP (Postgres or your DB of choice)
What it does: lets Claude run queries against your database (read-only by default — and you absolutely want to keep it that way).
Why it's essential: "Show me the top 10 customers by revenue last month" becomes an instant answer instead of a Slack thread to data team. For product/business people this is a quality-of-life transformation.
Set up a read-only Postgres user, point the MCP server at it, never let Claude write. Treat your prod DB like a glass museum case.
Security: this is not a toy
MCP servers run on your machine with your credentials. A misconfigured one could accidentally leak data or take destructive actions. Three rules:
- Tokens with minimum scope. If Claude only needs to read GitHub issues, give it a token that can read issues. Not a personal access token with full repo write.
- Read-only for databases. Always. There is no scenario where Claude should be running
DROP TABLEfrom a conversation. - Approve tool calls when starting. Claude Code asks before calling any MCP tool. Don't hit "always approve" until you understand what the tool does.
Where MCP fits with Skills and Agents
Three concepts often get confused. Here's the clean mental model:
Skills
Knowledge. Markdown files that teach Claude how to do specific tasks.
Agents
Workers. Specialized Claudes dispatched to do focused jobs end-to-end.
MCP servers
Tools. Connectors that let Claude take actions in real systems.
A complete workflow uses all three: an Agent that follows a Skill that calls MCP tools. Example: a "weekly summary" Agent (worker) that uses a "status report" Skill (knowledge) and the GitHub MCP (tools) to pull merged PRs from the week and write a summary.
What to do today
Pick two from the list and install them this afternoon. Filesystem + GitHub is the safest starter pair — high utility, low risk. Use Claude with them for a week. Notice which workflows got faster. Add a third server based on what you wished you could automate but couldn't.
And if you want the Agent layer built for you — already integrated with common MCP patterns, covering 9 categories of work — check out the Better with Claude Agents Pack. 107 production-ready Agents. Each one is a focused worker that composes well with Skills and MCP.