MCPs That Actually Help

There are over 500 MCP servers listed on GitHub. Most of them are junk. Experiments, toys, wrappers around things Claude Code already does natively. You don't need 500. You need maybe three or four.

The useful ones all pass a simple test: they give Claude Code access to live data it can't reach on its own. A production database. Your Sentry error feed. Open pull requests on GitHub. If Claude Code can already do it with a bash command, skip the MCP.

GitHub MCP

The single most useful MCP server, and it's not close. It lets Claude Code interact with GitHub directly: review pull requests, create issues, check assignments, read PR comments, search across repositories.

One command to install:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/

First time, you authenticate via OAuth. Run /mcp inside Claude Code to trigger it. After that, you can ask things like "show me all open PRs I need to review" or "create an issue for the login bug we discussed." Claude Code talks to the GitHub API directly instead of shelling out to gh commands.

Honest take: if you already use gh CLI comfortably, this is convenient but not life-changing for interactive work. Where it genuinely shines is autonomous tasks. Claude Code is working through a problem, realizes it needs to check a PR's status or read an issue description, and just... does it. No context switch. No copy-pasting URLs. That flow alone is worth the install.

Sentry MCP

If your team uses Sentry, install this immediately. It connects Claude Code to your Sentry project so it can pull up production errors, read stack traces, and check deployment health.

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

Same OAuth flow as GitHub. Authenticate with /mcp, then ask Claude Code to investigate production errors by pasting an error ID or just describing the symptoms. It pulls the full stack trace, event context, release info, and number of affected users. Then it traces back into your codebase and proposes a fix with the real error context right there.

Without the MCP, the workflow is: open Sentry in a browser tab, find the error, copy the stack trace, switch to your terminal, paste it in. With it, Claude gets structured access to the full event, not just a blob of stack trace text you pasted.

PostgreSQL and Database Access

Connecting Claude Code to a database sounds terrifying. For read-heavy workflows, though, it's ridiculously useful. The setup uses the Bytebase DBHub server:

claude mcp add --transport stdio db -- npx -y @bytebase/dbhub --dsn "postgresql://user:password@host:5432/mydb"

Now you can say "show me the 10 most recent orders with their customer names" and Claude Code writes the SQL, runs it, and shows you the results. It understands your schema, explores tables, builds queries from natural language.

The gotcha is output size. Database queries return a lot of data, and MCP responses exceeding 10,000 tokens trigger a warning. One careless SELECT * on a large table and you've flooded your context window with rows. Use a read-only replica if you're connecting to production. Keep queries focused. You can set MAX_MCP_OUTPUT_TOKENS as an environment variable to cap response sizes before they spiral.

Windows users: there's a second gotcha. Stdio-based servers using npx need to be wrapped with cmd /c because of how process spawning works. If the command above fails, wrap the npx call accordingly.

Environment Variables and Auth Patterns

Most MCP servers need API keys or tokens. Pass them as environment variables during setup:

claude mcp add --transport stdio myserver -- node server.js --env API_KEY=sk-xxx

Remote servers (HTTP transport like GitHub and Sentry) use OAuth 2.0 instead. The /mcp command handles the browser auth flow. Authenticate once, token gets stored for future sessions.

Watch out: if you put API keys in your MCP configuration, don't commit that file to version control. Your CLAUDE.md might reference MCP servers, but actual secrets belong in environment variables or a gitignored local config file.

Try It

Pick one server that matches a tool your team already uses. GitHub's the easiest starting point. If you query a PostgreSQL database regularly, try that one instead.

claude mcp add --transport http github https://api.githubcopilot.com/mcp/

After adding it, start a session and run /mcp to verify the connection. Then try something simple:

Show me my open pull requests and summarize which ones have failing checks.

If it works, you'll see Claude Code calling GitHub MCP tools instead of running gh commands. Structured API access, not shell commands.

What Not to Install

Every MCP server registers tools, and those tool descriptions eat context window space. Once you've got too many, Claude Code activates MCP Tool Search, which adds indirection: it searches for the right tool before calling it. It works, but it's slower than having a small, focused tool set.

Install servers for services you use daily. If you check Sentry once a week, don't leave that MCP running in every session. Add and remove servers as needed, or scope them to specific projects with project-level MCP config.

Security and Trust

Anthropic hasn't verified every third-party MCP server. When you install one, you're giving it access to your Claude Code session, potentially your filesystem, network, and credentials. That alone should give you pause.

The scarier risk is prompt injection. An MCP server that fetches content from the internet (web scraping, email, Slack messages) could return text with embedded instructions designed to manipulate Claude Code. If Claude acts on those instructions, you've got a serious problem.

Stick to servers maintained by the service provider themselves: GitHub, Sentry, Stripe. Be cautious with community servers that fetch arbitrary external content. Always review what a server actually does before installing it.

GitHub and Sentry MCPs are the best starting points. Database servers are powerful but watch output size. Fewer servers beats more servers. And /mcp is your go-to for auth and verifying connections.