The best MCP servers for Claude Code, and the ones to skip

How this list was cut

Four rules, applied hard. They are the reason this page is shorter than the lists it competes with.

• It must do something Claude Code cannot already do. That single test removes roughly half of every popular list.
• It must be maintained by the vendor whose system it touches, or be the obvious community standard. A third-party GitHub wrapper is a credential you handed to a stranger.
• Its tool surface must be small enough to describe. A server with ninety tools is a server the model will misuse.
• It must be installable in one command that we have verified against the vendor's own documentation.

claude mcp add --scope user --transport http github https://api.githubcopilot.com/mcp/x/all/readonly \
  --header "Authorization: Bearer $GITHUB_PAT"
claude mcp add --scope user --transport http context7 https://mcp.context7.com/mcp \
  --header "Authorization: Bearer $CONTEXT7_API_KEY"
claude mcp add --scope user playwright -- npx @playwright/mcp@latest

claude mcp list

The shortlist

The two everyone should have

GitHub

The one with the clearest payoff, because it collapses the loop of copying issue text and PR review comments into the terminal. GitHub runs the remote server itself, and it is unusual in exposing per-toolset endpoints so you can grant a narrow slice rather than everything.

# everything, read-only
claude mcp add --transport http github https://api.githubcopilot.com/mcp/x/all/readonly \
  --header "Authorization: Bearer YOUR_GITHUB_PAT"

# or just pull requests and issues, read-write
claude mcp add --transport http gh-pr https://api.githubcopilot.com/mcp/x/pull_requests \
  --header "Authorization: Bearer YOUR_GITHUB_PAT"

A browser server

Pick one, not both. Playwright is the better general driver: it exposes the page as a structured accessibility snapshot rather than screenshots, which means the model reads a tree instead of guessing at pixels. Chrome DevTools MCP is the better debugger, because it hands over the console, the network log, and real performance traces from a live Chrome.

# general automation and end-to-end checks
claude mcp add playwright -- npx @playwright/mcp@latest

# debugging a running page: console, network, performance
claude mcp add chrome-devtools --scope user -- npx chrome-devtools-mcp@latest

This is the category where an agent goes from writing plausible frontend code to verifying it. "Load the page, click through checkout, tell me what threw" is a genuinely different workflow from "read the component and guess".

Documentation and errors

Context7

The highest value-per-token server on this page. Models are confidently wrong about library APIs that changed after training, and the usual workaround is pasting documentation by hand. Context7 fetches version-specific docs and code examples for the library actually in your lockfile.

claude mcp add --transport http context7 https://mcp.context7.com/mcp \
  --header "Authorization: Bearer YOUR_API_KEY"

Sentry

Turns "the checkout page is broken" into a stack trace, a release, and a first-seen timestamp without leaving the terminal. It authenticates over OAuth, so no credential goes into your config file.

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

Issue trackers, docs, and design

# Linear (read-only variant available)
claude mcp add --transport http linear https://mcp.linear.app/mcp
claude mcp add --transport http linear-ro https://mcp.linear.app/mcp/readonly

# Atlassian: Jira, Confluence, JSM, Bitbucket
claude mcp add --transport http atlassian https://mcp.atlassian.com/v1/mcp/authv2

# Notion
claude mcp add --transport http notion https://mcp.notion.com/mcp

# Asana still only exposes SSE, which is deprecated but works
claude mcp add --transport sse asana https://mcp.asana.com/sse

Figma

Worth it only if you build UI against Figma files, in which case it is transformative: the agent reads the actual frame, variables, and spacing instead of inferring them from a screenshot you pasted.

claude mcp add --transport http figma https://mcp.figma.com/mcp

# or, as a plugin that bundles the server plus skills
claude plugin install figma@claude-plugins-official

Databases and infrastructure

The category with the highest upside and the sharpest edge. A schema-aware agent writes dramatically better migrations. A write-capable production credential attached to an agent is an incident with a date on it.

# any Postgres or MySQL, via DBHub
claude mcp add --transport stdio db -- npx -y @bytebase/dbhub \
  --dsn "postgresql://readonly:pass@localhost:5432/app_dev"

# Supabase, hosted; sign in at /mcp afterwards
claude mcp add --transport http supabase https://mcp.supabase.com/mcp

# or the local package, where --read-only and --project-ref are available
claude mcp add supabase-local -- npx -y @supabase/mcp-server-supabase \
  --read-only --project-ref=YOUR_PROJECT_REF

# Cloudflare ships sixteen scoped servers; add only the ones you use
claude mcp add --transport http cf-docs https://docs.mcp.cloudflare.com/mcp
claude mcp add --transport http cf-observability https://observability.mcp.cloudflare.com/mcp

The ones to skip

These appear near the top of most lists and are actively counterproductive in Claude Code, because the capability already exists natively and costs nothing.

Keeping the list honest

# what is configured, and is it actually connecting?
claude mcp list

# what did I add and forget
claude mcp get some-server

# remove it, from whichever scope it lives in
claude mcp remove some-server

# and clear any OAuth credential it left behind
claude mcp logout some-server

Claude Code defers MCP tool definitions by default now, which softens the cost of a long list but does not remove it: server instructions still load at session start, and tool output still enters your context permanently once a call returns. A server you have not used in a month is pure overhead plus a live credential.

Questions people ask

Keep reading