Claude Code
Best MCP Servers for Claude Code in 2026
By Ethan Halfhide · Updated October 20, 2026
BestMCPTools
Best MCP Servers for Claude Code in 2026
bestmcptools.org
Last verified 2026-10-20
The best MCP servers for Claude Code in 2026 are Context7 for current library docs, the official GitHub MCP Server for repos and pull requests, Playwright for browser automation, and Sentry for production errors, with Supabase, Chrome DevTools, Vercel and Exa for more specific jobs. Each documents a Claude Code install path, and six of the eight offer a hosted remote server. Below we rank all eight from the developer tools and neighbouring categories, with the install command for each and how scopes, .mcp.json and /mcp work.
TL;DR, Quick Verdict
Start with Context7 and the official GitHub MCP Server: they fix stale library knowledge and give Claude your repository work outside the local checkout. Add Playwright if you build a UI and Sentry once you have production traffic. Pick Supabase or Vercel only if you already use them, and Chrome DevTools when you need performance traces. Skip the Filesystem server, since Claude Code already edits files natively, and avoid the archived GitHub and PostgreSQL reference servers.
Quick Picks
Context7 MCP
Pulls version-specific docs and code examples into context through two tools.
GitHub MCP Server
GitHub-hosted remote endpoint with toolsets for repos, issues and pull requests.
Playwright MCP
Drives a real browser through accessibility snapshots, no vision model needed.
Sentry MCP
Remote OAuth server that can be scoped to one Sentry organization or project.
Supabase MCP
Hosted server with project scoping and a read-only mode in the URL.
Chrome DevTools MCP
Records traces, reads network requests and console messages in live Chrome.
How we chose
We compared official information only: each project's README, vendor docs, the Claude Code MCP reference and pricing pages, all read on October 20, 2026. We did not test these servers. A server qualified if its product owner maintains it, it documents a Claude Code install path, and it adds something Claude Code's built-in tools lack. We ranked by how many coding projects benefit, then by how little setup and permission it needs.
How to add a server with claude mcp add
The Claude Code docs describe three shapes. A remote server is added with claude mcp add --transport http <name> <url>; Claude Code tries HTTP first and falls back to SSE, which the docs mark as deprecated. A local stdio server takes the server name and Claude Code's own flags (--transport, --env, --scope) before a double dash, then the launch command after it, as in claude mcp add example --env API_KEY=your-key -- npx -y @example/mcp-server. A JSON block from another client goes in with claude mcp add-json. Afterwards, claude mcp list and claude mcp get <name> show status, claude mcp remove deletes a server, and /mcp inside a session shows connections and handles OAuth sign-in. Our list of MCP servers that work in Cursor covers that editor's side.
Scopes and .mcp.json for teams
Servers live in one of three scopes. Local, the default, loads only in the current project and is stored privately in ~/.claude.json. User scope (--scope user) loads in all your projects, which suits Context7 or Chrome DevTools. Project scope (--scope project) writes .mcp.json at the repository root so the team shares it through version control, and Claude Code asks for approval before using those servers interactively. If one name exists in several scopes, local beats project, which beats user. Keep secrets out of the committed file with ${VAR} or ${VAR:-default} expansion, which works in command, args, env, url and headers.
Which one fits your project
For almost any codebase, Context7 MCP server comes first because it answers current API questions that training data gets wrong. If your code is on GitHub, the GitHub MCP Server handles issues and pull requests, with an optional Actions toolset. Front-end teams should add Playwright MCP server for flows, or Chrome DevTools when a page is slow. Teams with live users get the most from Sentry MCP server. Supabase and Vercel only make sense if you already use those platforms. Exa is niche because Claude Code already ships WebSearch and WebFetch.
Local vs remote setup
Six picks offer a hosted remote server: Context7 (mcp.context7.com/mcp), GitHub (api.githubcopilot.com/mcp), Sentry (mcp.sentry.dev/mcp), Supabase (mcp.supabase.com/mcp), Vercel (mcp.vercel.com) and Exa (mcp.exa.ai/mcp). Remote servers need no local runtime and usually sign in through OAuth from /mcp, and Claude Code reconnects one that drops mid-session. Playwright and Chrome DevTools run locally because they control a browser on your machine, and the docs note stdio servers are not reconnected automatically. GitHub also offers a local Docker image or binary, for example for GitHub Enterprise Server.
Costs and accounts
As of October 2026, none of the vendor pages we read lists a separate charge for the MCP server itself; costs come from the service behind it. Playwright and Chrome DevTools are open source with no account. Context7 recommends a free API key for higher rate limits. Sentry's Developer plan is $0 for one user and lists MCP access. Supabase's Free plan is $0 and Pro starts at $25 a month. Vercel MCP is available on all plans, with Hobby at $0 for personal, non-commercial use and Pro at $20 a month. Exa allows keyless, rate-limited use, its Starter plan includes $10 of free credits a month, and search costs from $4 per 1,000 requests. GitHub needs an account plus a token or OAuth, and any applicable GitHub policies enabled.
Security and permissions
The Claude Code docs say to verify you trust each server before connecting it, because servers that fetch external content can expose you to prompt injection. Scope narrowly: Sentry and Supabase accept URLs scoped to one organization or project, and Supabase's read_only=true runs queries as a read-only Postgres user. Limit tools: GitHub's default toolsets are context, repos, issues, pull_requests and users, and its local build supports --read-only. Keep tokens out of git with ${VAR} expansion and minimum-scope tokens, as GitHub recommends. Chrome DevTools sends usage statistics to Google by default; --no-usage-statistics opts out.
Why Filesystem and the old reference servers are not on the list
Claude Code ships built-in Read, Write, Edit and Bash tools, and extra folders can be added with --add-dir or additionalDirectories, so the Filesystem MCP server duplicates what you have. The old GitHub and PostgreSQL MCP server reference servers sit in the modelcontextprotocol/servers-archived repository, which says it is no longer maintained and offers no security guarantees; our GitHub entry describes the maintained github/github-mcp-server instead. For picks across all clients, see our list of the best MCP servers.
What we could not verify
Context7's pricing page did not render readably, so we could not confirm paid plan prices or exact free rate limits. GitHub's README lists no price for the hosted remote server. We found no separate usage caps for Sentry MCP calls. Exa documents a Claude Code plugin rather than a claude mcp add command, so we show the plugin command only.
Quick Comparison
| Server | Best for | Transport | Auth | Maintainer | Price (as of October 2026) |
|---|---|---|---|---|---|
| Context7 MCP | Current library docs | Remote HTTP or local stdio | API key (recommended) | Upstash | Free API key; paid tiers not verified |
| GitHub MCP Server | Repos, issues, pull requests | Remote HTTP or local Docker/binary | PAT or OAuth | GitHub | Needs GitHub account; no MCP price listed |
| Playwright MCP | Browser automation | Local stdio | None | Microsoft | Free, open source |
| Sentry MCP | Production errors and traces | Remote HTTP (stdio for self-hosted) | OAuth | Sentry | Developer plan $0 includes MCP access |
| Supabase MCP | Postgres, auth, edge functions | Remote HTTP | OAuth or access token | Supabase | Free plan $0; Pro from $25/month |
| Chrome DevTools MCP | Performance and debugging | Local stdio | None | Google Chrome DevTools | Free, open source |
| Vercel MCP | Deployments, logs, docs | Remote HTTP | OAuth | Vercel | All plans; Hobby $0, Pro $20/month |
| Exa MCP | Web search and research | Remote HTTP or local npm | Keyless, OAuth or API key | Exa | Starter free ($10 credits/month); search from $4 per 1k |
The Tools, Ranked
Context7, an Upstash project, fetches up-to-date, version-specific documentation and code examples into Claude's context through two tools: resolve-library-id, which maps a library name to a Context7 ID, and query-docs, which returns docs for that ID. You can name a library directly, such as /supabase/supabase, or mention a version in your prompt. The README suggests a CLAUDE.md rule so Claude uses it automatically, and warns that indexed projects are community-contributed.
Pros
- ✓Remote URL means no local install
- ✓Only two tools, so a small context footprint
- ✓Works at user scope across every project
Cons
- ✗Free API key recommended for higher rate limits
- ✗Library content is community-contributed
- ✗Backend components are private, only the MCP server is open source
Quick Install
claude mcp add --scope user --header "Authorization: Bearer YOUR_API_KEY" --transport http context7 https://mcp.context7.com/mcp
Example
Ask Claude to build a Next.js proxy that checks a JWT cookie and have it pull the current Next.js docs first.
Use GitHub's official server from github/github-mcp-server, not the older @modelcontextprotocol/server-github package, which is archived with a notice that development moved to GitHub. The hosted endpoint is api.githubcopilot.com/mcp, and GitHub's Claude Code guide adds it with a personal access token header (Claude Code 2.1.0 or earlier uses claude mcp add --transport http instead). Default toolsets are context, repos, issues, pull_requests and users. The local Docker image or binary can sign in with OAuth instead of a token, supports --read-only mode and works with GitHub Enterprise Server, which the remote server does not support.
Pros
- ✓Maintained by GitHub with a hosted endpoint
- ✓Toolsets let you expose only what you need
- ✓Local build supports --read-only mode
Cons
- ✗Token setup needed for the remote server in Claude Code
- ✗Applicable GitHub policies may need enabling first
- ✗Our listing still links the archived reference server
Quick Install
claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'Example
Have Claude read a bug issue, push a fix to a new branch and open a pull request that references it.
Microsoft's Playwright MCP lets Claude control a browser through structured accessibility snapshots instead of screenshots, so no vision model is needed. It needs Node.js 18 or newer and runs locally, headed by default, with --headless and --isolated flags and optional vision, pdf and devtools capabilities via --caps. The README notes that coding agents may find the Playwright CLI with skills more token-efficient, while MCP suits loops that need persistent browser state.
Pros
- ✓Official Claude Code command in the README
- ✓Accessibility-tree approach, no screenshots required
- ✓Many flags for headless, device emulation and isolation
Cons
- ✗Snapshots can be large in context
- ✗Persistent profile cannot be shared by parallel clients
- ✗Microsoft suggests the CLI may suit coding agents better
Quick Install
claude mcp add playwright npx @playwright/mcp@latest
Example
Ask Claude to open your local dev server, sign up with a sample account and report any step that fails.
Sentry's MCP server is a remote service at mcp.sentry.dev, built for human-in-the-loop coding agents such as Claude Code. All connections use OAuth, and the URL can be scoped to one organization or project. Sentry also ships a Claude Code plugin with a sentry-mcp subagent, plus a stdio mode for self-hosted Sentry. Its AI-powered search tools need an LLM provider configured.
Pros
- ✓OAuth, no tokens to paste
- ✓URL scoping to one org or project
- ✓Plugin option with a dedicated subagent
Cons
- ✗Natural-language search tools need an LLM provider
- ✗Stdio mode is described as a work in progress
- ✗Only useful once you send events to Sentry
Quick Install
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/{organizationSlug}/{projectSlug}Example
Paste a Sentry issue link and ask Claude to find the stack trace, locate the code and propose a fix.
Supabase's hosted server at mcp.supabase.com/mcp prompts you to log in, then lets Claude list tables, run SQL, apply migrations, deploy edge functions and search Supabase docs. Its Claude Code command uses project scope, so the config lands in .mcp.json. URL parameters handle safety: project_ref limits access to one project, read_only=true runs queries as a read-only Postgres user, and features restricts tool groups. Supabase recommends branching for safer schema changes.
Pros
- ✓Official Claude Code command at project scope
- ✓Read-only and project scoping in the URL
- ✓Local CLI endpoint for Supabase local development
Cons
- ✗Write tools can change a live database
- ✗Self-hosted and CLI modes expose a limited tool subset
- ✗Only relevant if you use Supabase
Quick Install
claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp?features=docs%2Caccount%2Cdatabase%2Cdebugging%2Cdevelopment%2Cfunctions%2Cbranching"
Example
Ask Claude to inspect the schema in read-only mode and draft a migration for a new orders table.
Chrome DevTools MCP, published by the ChromeDevTools organization on GitHub, gives Claude a live Chrome instance to control and inspect. It records performance traces and extracts insights, analyses network requests, takes screenshots and reads console messages with source-mapped stack traces, using Puppeteer for automation. It needs a Node.js LTS release and current stable Chrome. Usage statistics go to Google by default, and performance tools may send trace URLs to the CrUX API unless you pass --no-performance-crux.
Pros
- ✓Real DevTools traces, not just clicks
- ✓Slim and headless modes for basic tasks
- ✓Official Claude Code command and plugin
Cons
- ✗Usage statistics on by default
- ✗Officially supports Chrome only
- ✗Exposes browser data to the MCP client
Quick Install
claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest
Example
Ask Claude to record a trace of your landing page and explain the largest contentful paint delay.
Vercel MCP is Vercel's official remote server at mcp.vercel.com for searching Vercel docs, managing projects, inspecting deployments and querying Web Analytics. Vercel lists Claude Code as a supported client: add it with one command, then run /mcp to authorize with OAuth. Vercel only allows clients it has approved and asks for explicit consent for each client connection.
Pros
- ✓Official Claude Code instructions
- ✓OAuth with per-client consent
- ✓Available on all Vercel plans
Cons
- ✗Can manage your teams, projects and deployments
- ✗Only approved clients can connect
- ✗Useful only for Vercel-hosted projects
Quick Install
claude mcp add --transport http vercel https://mcp.vercel.com
Example
Ask Claude why the latest preview deployment failed and have it read the build logs.
Exa MCP
Exa's hosted server at mcp.exa.ai/mcp offers web_search_exa and web_fetch_exa by default, with web_search_advanced_exa and agent_run available through a tools parameter. No API key is needed to start; OAuth or a key unlocks agent runs. For Claude Code, Exa documents a plugin that bundles the MCP server with search skills. Add it only if you want Exa's filters or agent runs on top of Claude Code's built-in web tools.
Pros
- ✓Keyless start, no sign-up
- ✓Advanced filters for domains and dates
- ✓Plugin bundles MCP and skills
Cons
- ✗Overlaps with built-in WebSearch and WebFetch
- ✗Agent runs are usage-based
- ✗Setting a tools list replaces the defaults
Quick Install
claude plugin install exa@claude-plugins-official
Example
Ask Claude to find recent changelog posts for a library and summarise breaking changes.
The Verdict
For most Claude Code users the right starting set is small: Context7 MCP server at user scope for current docs and the official GitHub MCP Server for repository work. Add a browser server when you need one, and Sentry once real users find your bugs.
Treat Supabase and Vercel as platform add-ons, scope them to one project and prefer read-only mode at first. Tool search keeps context cost low, but every server is another set of permissions, so add servers one at a time, check them in /mcp and remove what you stop using. Browse more in browser automation.
Frequently Asked Questions
How do I add an MCP server to Claude Code?
Run claude mcp add in your terminal. For a hosted server use claude mcp add --transport http <name> <url>. For a local server put the launch command after a double dash, for example claude mcp add example --env API_KEY=your-key -- npx -y @example/mcp-server. Then type /mcp inside Claude Code to check status or sign in with OAuth.
How many MCP servers should I run in Claude Code?
There is no fixed limit. Claude Code's tool search is on by default and loads only tool names and server instructions at session start, so extra servers have little context cost. The practical limit is your context window and how many permissions you are comfortable granting, so most people keep a short list of servers they use weekly.
Do I need the Filesystem MCP server with Claude Code?
Usually not. Claude Code has built-in Read, Write, Edit and Bash tools, and you can grant extra folders with --add-dir or the additionalDirectories setting.
Is the GitHub MCP server archived?
The old reference package, @modelcontextprotocol/server-github, is archived and no longer maintained. GitHub now maintains the official server in the github/github-mcp-server repository, with a hosted endpoint at api.githubcopilot.com/mcp and a local Docker image.
What is the difference between local, project and user scope?
Local is the default and loads the server only in the current project, stored privately in ~/.claude.json. Project scope writes .mcp.json in the repository so teammates share it. User scope loads the server in all your projects. If one name exists in several scopes, local wins, then project, then user.
Sources
- Claude Code docs: Connect Claude Code to tools via MCP
- Claude Code docs: Tools reference
- Context7 GitHub repository
- Context7 docs: MCP clients
- GitHub MCP Server repository
- GitHub MCP Server: Install in Claude
- Archived MCP reference servers
- Playwright MCP repository
- Sentry MCP server
- Sentry MCP repository
- Sentry pricing
- Supabase docs: Model context protocol
- Supabase pricing
- Chrome DevTools MCP repository
- Chrome DevTools MCP client configurations
- Vercel docs: Vercel MCP
- Vercel pricing
- Exa docs: Exa MCP
- Exa MCP repository
- Exa pricing
About the author
Ethan Halfhide
Ethan Halfhide is an AI technologist who works with AI agents and Model Context Protocol integrations. He curates BestMCPTools' server directory, checking each server's documentation, supported clients and setup steps.