The AI Loadout

Best MCP Servers for Claude Code (Nov 2026)

Published · Updated

By The AI Loadout editors · How we pick · Nichelogix, LLC

Start with three: Context7 for current library docs, GitHub's official server for issues and PRs, and Playwright or Chrome DevTools so Claude can check its own work in a real browser. Add Supabase or DBHub for data, Sentry for production errors and Firecrawl for web research. All eight have a free way in.

Security pick

HOL Guard helps check agent skills, MCP servers and plugins before you install them. On supported integrations, it evaluates risky actions locally and can allow, ask or block before they run.

Try HOL GuardAffiliate link. We earn a commission if you buy, at no cost to you.

Use code GUARD25-CA564976 for 25% off your first payment.

For agent builders: get verified and listed on HOL

Blue adds a checkmark to one listing. Gold covers every listing you own and adds priority re-scans. Featured puts your plugin or extension in HOL directory placements.

  • Get HOL Blue Affiliate link. We earn a commission if you buy, at no cost to you. Code: BLUE25-CA564976 for 25% off your first payment.
  • Get HOL Gold Affiliate link. We earn a commission if you buy, at no cost to you. Code: GOLD25-CA564976 for 25% off your first payment.
  • Get Featured placement Affiliate link. We earn a commission if you buy, at no cost to you. Code: FEATURED10-CA564976 for 10% off your first payment.

The 8 MCP servers worth adding

We picked servers that are maintained by the company behind the product (or by a well known open source team), that install with one claude mcp add line, and that give Claude something it cannot do with its built-in tools. Claude Code can already read, edit and search files, run shell commands and fetch web pages on its own (tools reference), so a server has to add live access to something outside your repo to earn a slot.

Prices were checked on Nov 2, 2026 on each vendor's own pricing page. None of these picks are sponsored and none of the links pay us.

  1. Context7 (Free 1,000 calls/mo, Pro $10/seat): Current, version-specific library docs so Claude stops guessing at APIs.
  2. GitHub MCP Server (Free, MIT): Issues, pull requests, Actions runs and security alerts, with a read-only URL.
  3. Playwright MCP (Free, Apache-2.0): Drives a real browser through accessibility snapshots to test flows.
  4. Chrome DevTools MCP (Free, Apache-2.0): Performance traces, console and network debugging, and Lighthouse audits.
  5. Supabase MCP (Free plan, Pro from $25/mo): Tables, SQL, logs and advisors for Supabase projects, with read-only mode.
  6. DBHub (Free, MIT): One lightweight server for Postgres, MySQL, SQL Server, MariaDB, Oracle and SQLite.
  7. Sentry MCP (Developer plan $0, Team $26/mo): Stack traces and issues from production errors, scoped to one project.
  8. Firecrawl MCP (Free 1,000 credits/mo, Hobby $19/mo): Search, scrape, map and crawl the web into clean markdown for research.

The loadout for each task

Most "best MCP servers" lists stop at the servers. A server only gives Claude access. The skill is what tells it how you want the job done, and the model and persona decide how carefully it works. Here is what we would run for the three jobs people ask us about most.

MCP servers, skills, model and persona per task. Checked Nov 2, 2026.
TaskMCP serversSkillsModelPersona
Web app developmentContext7, GitHub, Playwright or Chrome DevTools, SentryTDD skill, systematic debuggingClaude Sonnet 5.5, Opus 5.5 for big refactorsPragmatic Staff Engineer
SEO and contentFirecrawl, Chrome DevTools (Lighthouse audits)A voice skill you write once, plus the skills that ship with the Chrome DevTools pluginClaude Sonnet 5.5Fact-First Copy Chief
Data and analyticsDBHub (any SQL database) or Supabase, both read-onlyA short schema notes skill: table meanings, join keys, which numbers finance trustsClaude Opus 5.5 for tricky queriesHigh-Throughput DBA

The full coding setup, with editor and model alternatives, is on our best AI coding agent setup page. For the content side, see AI setup for SEO blog posts, and for data work, the best AI setup for data analysis.

Context7: stop Claude guessing at library APIs

Models learn a library once, at training time. Context7 fetches current, version-specific docs and code examples when Claude needs them, which cuts down on calls to functions that were renamed two releases ago. Mention the version in your prompt ("Next.js 16") and it matches it.

Price: the Free plan includes 1,000 API calls a month, and once you hit it you get 20 bonus calls a day until the month resets. Pro is $10 per seat a month with 2,000 calls per seat and private repos (Context7 plans). A free API key from the dashboard raises your rate limits.

Add Context7 (all projects)
claude mcp add --scope user --header "Authorization: Bearer YOUR_API_KEY" --transport http context7 https://mcp.context7.com/mcp

Context7 also offers a CLI plus skill mode with no MCP server at all (npx ctx7 setup lets you choose). If you already keep tool count low, that mode is worth a look.

GitHub MCP: issues, PRs and failed Actions runs

Claude Code can already run git and the gh CLI. GitHub's official server earns its place for the things that are clumsy from a shell: reading a failed Actions run, triaging a pile of issues, reviewing a PR with its comments, or checking Dependabot and code scanning alerts. It is free and open source (MIT). Organization policies on personal access tokens can block it, so check with your admin on a work account.

Turn on only what you need. The remote server takes a /readonly suffix on its URL, and GitHub's docs list toolset URLs such as issues or actions if you want fewer tools.

Add GitHub (remote, personal access token)
claude mcp add --transport http github https://api.githubcopilot.com/mcp/ --header "Authorization: Bearer YOUR_GITHUB_PAT"
Read-only variant
claude mcp add --transport http github https://api.githubcopilot.com/mcp/readonly --header "Authorization: Bearer YOUR_GITHUB_PAT"

Step by step with token scopes: connect Claude Code to GitHub. Our notes on the repo: github-mcp-server.

Playwright or Chrome DevTools: let Claude look at what it built

This is the server that changes results the most for front end work. Without a browser, Claude writes a component and tells you it works. With one, it opens the page, clicks through the flow and reads the console errors itself.

Pick one per project, not both. Playwright MCP from Microsoft drives the page through accessibility snapshots instead of screenshots, which suits clicking through forms and flows. Chrome DevTools MCP from Google adds performance traces, network requests, source-mapped console errors and a lighthouse_audit tool that scores accessibility, SEO and best practices. Both are free (Apache-2.0).

Add Playwright MCP
claude mcp add playwright npx @playwright/mcp@latest
Add Chrome DevTools MCP (all projects)
claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

Two details the other lists skip. Microsoft's own README says coding agents may do better with the Playwright CLI plus skills, because CLI calls avoid loading large tool schemas and accessibility trees into context. And Chrome DevTools MCP sends usage statistics to Google by default and may send trace URLs to the CrUX API; the flags --no-usage-statistics and --no-performance-crux turn those off.

Chrome DevTools also ships as a Claude Code plugin that installs the server plus its skills: /plugin marketplace add ChromeDevTools/chrome-devtools-mcp, then /plugin install chrome-devtools-mcp@chrome-devtools-plugins. More on the repo: playwright-mcp.

Supabase or DBHub: query the real database, read-only

If your app runs on Supabase, use the official server. It signs in with OAuth (no token to paste) and can list tables, run SQL, read logs and run Supabase's security and performance advisors. Add read_only=true to run every query as a read-only Postgres user, and project_ref to limit it to one project. Supabase's Free plan covers 2 active projects with a 500 MB database; Pro starts at $25 a month (Supabase pricing).

For anything else (Postgres, MySQL, SQL Server, MariaDB, Oracle, SQLite) use DBHub from Bytebase. It is free (MIT), loads 2 tools by default, and is the example Anthropic uses in the Claude Code docs. Connect it with a database user that can only read.

Add Supabase, scoped and read-only
claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp?project_ref=YOUR_PROJECT_REF&read_only=true"
Add DBHub for Postgres
claude mcp add --transport stdio db -- npx -y @bytebase/dbhub --dsn "postgresql://readonly:pass@host:5432/db"

After adding Supabase, run /mcp in a regular terminal session, pick supabase and authenticate. The longer walkthrough is in query a database with AI.

Sentry: fix the bug users actually hit

Paste a Sentry issue link and Claude can pull the stack trace, the breadcrumbs and the trace data, then go find the code. Sentry's hosted server uses OAuth, and putting your organization and project in the URL keeps Claude on the right project. The free Developer plan is one user and lists MCP access; Team is $26 a month billed annually (Sentry pricing).

Add Sentry, scoped to one project
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/YOUR_ORG/YOUR_PROJECT

Then run /mcp to sign in. Sentry also publishes a Claude Code plugin (claude plugin marketplace add getsentry/sentry-mcp, then claude plugin install sentry-mcp@sentry-mcp) that adds a Sentry subagent Claude hands error questions to.

Firecrawl: web research that comes back as clean markdown

Claude Code's built-in fetch is fine for one page. For SEO work you usually want more: search results with the pages scraped, a map of every URL on a competitor's site, or a crawl of a docs section. Firecrawl does that and returns markdown.

Price: the Free plan gives 1,000 credits a month (about 1,000 scraped pages or 500 searches) with no card. Hobby is $19 a month, or $16 a month billed yearly, for 5,000 credits (Firecrawl pricingAffiliate link. We earn a commission if you buy, at no cost to you.). There is also a keyless endpoint, https://mcp.firecrawl.dev/v2/mcp, that allows rate limited scrape, search and parse without an account.

Add Firecrawl (browser sign-in)
claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp-oauth

Complete sign-in with /mcp. Our Firecrawl tool page compares it with Tavily, and scrape a website with AI has the step by step.

A .mcp.json you can commit to your repo

Add a server with --scope project and Claude Code writes it to .mcp.json at the root of your repo, so everyone who clones the project gets the same servers. Claude Code asks each person to approve project servers before it uses them.

The trick is keeping secrets out of the file. Claude Code expands ${VAR} and ${VAR:-default} in command, args, env, url and headers, so the file holds the variable name and each developer keeps the value in their own shell. Here is a web app starter:

  • A remote entry needs "type": "http". A bare url with no type is a config error in Claude Code.
  • Inside a remote server's url and headers, Claude Code reads its own and cloud provider credential variables (like ANTHROPIC_API_KEY) as empty, so a shared file cannot ship those keys to a server. Use your own variable names such as GITHUB_PAT.
  • If a variable is unset and has no default, the config still loads and claude mcp list warns you about it.
  • Changed your mind about a server you approved? claude mcp reset-project-choices clears those approvals.
.mcp.json (repo root)
{
  "mcpServers": {
    "context7": {
      "type": "http",
      "url": "https://mcp.context7.com/mcp",
      "headers": { "Authorization": "Bearer ${CONTEXT7_API_KEY}" }
    },
    "github": {
      "type": "http",
      "url": "https://api.githubcopilot.com/mcp/readonly",
      "headers": { "Authorization": "Bearer ${GITHUB_PAT}" }
    },
    "playwright": {
      "type": "stdio",
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    },
    "supabase": {
      "type": "http",
      "url": "https://mcp.supabase.com/mcp?project_ref=${SUPABASE_PROJECT_REF}&read_only=true"
    }
  }
}

MCP server or skill? The rule we use

Use an MCP server when Claude needs to reach something that lives outside your repo and changes: a database, an issue tracker, an error feed, a live web page. Use a skill when Claude needs to know how you want a job done: the steps, the checklist, the house style, which command to run. Anthropic's docs suggest a skill when you keep pasting the same instructions or procedure into chat.

The two work best together. The Supabase server gives Claude your schema; a 30 line schema notes skill tells it that orders.total is in cents and refunds live in a separate table. And when a vendor ships a good CLI, a skill that calls the CLI is often lighter than a server. Context7, Playwright and Chrome DevTools all now offer that route next to their MCP servers.

New to skills? Start with how to install Claude Code skills and the best Claude Code skills for web development.

Do more servers slow Claude down?

Less than they used to. Claude Code's tool search is on by default: only tool names and server instructions load when a session starts, and full tool definitions load when Claude needs them. Anthropic's docs say that means adding more servers has minimal impact on your context window.

Tool output is a different story. Claude Code warns when a single MCP tool result passes 10,000 tokens and caps results at 25,000 tokens by default (MAX_MCP_OUTPUT_TOKENS changes it). A full page scrape or a select * can hit that, so ask for the columns or sections you need. Running /doctor lists unused MCP servers next to their context cost, which makes pruning easy.

Keep the blast radius small

Every server runs with your permissions. Anthropic's own docs warn that servers which fetch outside content can expose you to prompt injection, and Supabase's docs walk through an example where a support ticket tells the model to dump a table. A few habits cover most of the risk:

  • Read-only first: read_only=true on Supabase, a read-only database user for DBHub, the /readonly GitHub URL.
  • Scope to one project: project_ref on Supabase, org and project in the Sentry URL.
  • Keep tool approvals on for anything that can write, especially in sessions that also browse or scrape the web.
  • Never put a real token in .mcp.json. Use ${VAR} and keep the value in your shell or a gitignored .env.
  • Don't add the same server twice. Claude Code treats two entries with the same URL or command as duplicates and keeps one, so an old copy in another scope can hide the one you just edited.

What to skip

A filesystem server, unless you need folders outside the project: Claude Code already reads, writes and searches files. Random community wrappers for a product that has an official server: Supabase, Sentry, GitHub and Firecrawl all run their own now. And the old reference Postgres server from the MCP repo, which has moved out of the main servers repo; DBHub is the one Anthropic's docs point to.

What this setup costs per month

Claude Code itself needs a paid Claude plan. Pro is $20 a month billed monthly, or $17 a month on the annual plan, and Claude Code is included in all paid plans (Claude pricing). Every server above has a free tier or is free software, so a solo developer can run the whole web dev loadout for $20 a month. The first paid upgrades people usually hit are Firecrawl Hobby ($19) for heavier crawling and Context7 Pro ($10) for private repos.

Set it up in ten minutes

Add the servers with the commands above, then check them:

Check what is connected
claude mcp list
claude mcp get github

Inside a session, /mcp shows status and handles sign-in for OAuth servers. If you want ready-made skills to pair with these servers, our own Solo Founder Pack ($9) bundles seven of them. The general walkthrough for any agent is how to use MCP servers.

Ready skill packs

The AI Loadout sells ready skill packs you can install with one terminal command after purchase. No refunds (all sales final) because packs are downloadable. If install fails, support@theailoadout.com will help fix it.

Browse skill packs

FAQ

How do I set up MCP servers in Claude Code?

Run claude mcp add in your terminal. For a hosted server use claude mcp add --transport http <name> <url>, adding --header for an API key. For a local one use claude mcp add <name> -- <command> [args]. Then run claude mcp list to confirm, and /mcp inside a session to sign in to OAuth servers.

What is the difference between MCP servers and skills in Claude Code?

An MCP server gives Claude live access to an outside system, such as a database, GitHub or a browser. A skill is a SKILL.md file of instructions that tells Claude how to do a job your way. Use a server for access and a skill for know-how. Most good setups use both.

What does a .mcp.json example look like?

It is a JSON file at the repo root with a mcpServers object. Remote servers need "type": "http" and a url; local servers use command and args. Reference secrets as ${VAR} instead of pasting them, so the file is safe to commit. The full example on this page sets up Context7, GitHub, Playwright and Supabase.

Where does Claude Code store MCP server settings?

Local scope (the default) and user scope are stored in ~/.claude.json. Local applies to the current project only, user applies to all your projects. Project scope writes .mcp.json in the repo root so the whole team shares it, and Claude Code asks each person to approve those servers.

Do MCP servers fill up Claude Code's context window?

Tool definitions mostly do not anymore. Tool search is on by default, so only tool names and server instructions load at the start and full definitions load when needed. Large tool results still cost context; Claude Code warns above 10,000 tokens and caps results at 25,000 by default.

Which MCP servers are free?

Playwright MCP, Chrome DevTools MCP, GitHub's MCP server and DBHub are free open source software. Context7 (1,000 calls a month), Supabase (2 active projects), Sentry (Developer plan, one user) and Firecrawl (1,000 credits a month) have free tiers. Prices checked Nov 2, 2026.

Are any of these picks sponsored?

No. None of the eight servers paid for placement, and the order is our own. Two links can earn us a commission: the Firecrawl pricing link and the HOL Guard box. Each is labeled as an affiliate link where it appears. Our own Solo Founder Pack is labeled as ours.

Sources

Disclosure: product links on this page go to The AI Loadout packs we sell. Affiliate and sponsored links elsewhere on the site are always labeled and do not change our picks. Read our methodology.

Related