Best MCP servers for AI agents in 2026: 10 picks compared

Which MCP servers should your AI agent use?

The best MCP servers for AI agents in 2026 are AgentMail for email, Firecrawl for web research, GitHub for code, Notion for team knowledge, Playwright for browser automation, Context7 for library docs, Supabase for databases, Linear for issue tracking, Ahrefs for SEO data, and Stripe for billing. Each one gives an agent a capability it can't fake with a generic tool.

Most agents that do real work need a few of these at once. An agent sourcing suppliers has to read their websites, email them for quotes, and log the answers where the team can see them. That's three servers for one job.

Below is what each server does, what to check before you connect it, and the command to add it to Claude Code.

Every server here is built and maintained by the company behind the product, and we checked each one against its official docs on September 30, 2026. The order is our recommendation by use case, not a usage count or a benchmark. AgentMail is our product, so weigh the number one spot with that in mind. We wrote this guide as part of a content partnership with Firecrawl.

1. AgentMail: best for email and agent-owned inboxes

AgentMail gives an AI agent an email address of its own, so it can talk to customers, vendors, and other agents without borrowing somebody's Gmail. Each inbox keeps the full history of its conversations.

The first email is the easy part. Then the vendor replies with a question, a customer sends a PDF, or someone moves the deadline. AgentMail groups related messages into threads, so the agent reads the whole exchange before it decides what to send next.

The MCP server exposes 36 tools covering inboxes, threads, messages, drafts, attachments, and allow and block lists. It supports OAuth, which AgentMail recommends for Claude products, and an API key header for clients like Cursor and Windsurf. If the agent should wake up when new mail arrives, point AgentMail webhooks at your application. MCP handles the mailbox, and your app decides when the agent runs.

Best for agents that need to send, receive, and reply to email from an inbox of their own.

Docs: AgentMail MCP

claude mcp add --transport http agentmail https://mcp.agentmail.to/mcp

2. Firecrawl: best for web research and data extraction

Firecrawl is the best MCP server for agents that need live, structured context from the web. Your agent can search for sources, open the pages that matter, and get back clean Markdown or structured data, even when the page is JavaScript-heavy, a PDF, or behind a login. It works in any MCP client.

Say your agent is comparing vendors. Here's what it can reach for along the way:

• Firecrawl Search finds the right sources on the live web. It's one of the most token-efficient search APIs for coding agents in 2026.
• Firecrawl Scrape reads a page and, given a JSON schema, returns only the fields you asked for, such as names, prices and source links.
• Firecrawl Map and Firecrawl Crawl find and read every relevant page when the details are spread across a site.
• Firecrawl Parse turns PDFs and other documents into text.
• Firecrawl Interact opens a page in a real browser to get past clicks, forms and logins.
• Firecrawl Agent takes the whole question and searches across sources on its own when you don't know where to look.
• Firecrawl Monitor lets you know when a page changes, such as a vendor updating its pricing.

Some questions are better answered from a dedicated source than from a general web search, so Firecrawl also includes:

• Research Index for papers from PubMed, bioRxiv, medRxiv and arXiv.
• Developer Index for GitHub issues, pull requests, READMEs and docs. It ranked first in a recent developer retrieval benchmark, ahead of tools like Context7, Exa and Parallel.
• Alexandria, Firecrawl's catalogue of data providers, for structured data like filings, public records, listings, quotes and job posts.

Best for agents that need fresh, structured context from the web, including sites they've never seen before.

Docs: Firecrawl MCP, tool reference, Alexandria

claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp-oauth

For keyless access, use https://mcp.firecrawl.dev/v2/mcp instead.

3. GitHub: best for repository workflows

GitHub's official server lets an agent work with repositories, issues, pull requests, and Actions. An agent chasing a bug can read the issue, open the relevant files, and put up a pull request with the fix.

You choose what the agent can reach through toolsets such as repos, issues, pull_requests, actions, and code_security. There's a read-only mode that drops every write tool, and we'd turn it on for any agent whose job is investigation. The server runs remotely at GitHub or locally in Docker, and authenticates with OAuth or a personal access token.

Best for agents that need to triage issues, review pull requests, or ship fixes in a repo.

Docs: GitHub MCP Server

claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'

4. Notion: best for workspace and database automation

Notion MCP connects an agent to the docs and databases your team already keeps in Notion. It's where a lot of agents should write down what they found: research notes, a vendor database, a summary of a long email thread.

The hosted server uses OAuth and can only see what the connected user can see. Agents can search Notion and connected sources, read and edit pages, create pages and databases, and upload files up to 20 MiB. Some search options depend on your Notion plan, so check the supported tools list before you rely on them.

Best for agents that need to read and update the docs and databases a team already works in.

Docs: Notion MCP

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

Run /mcp inside Claude Code afterward to finish the OAuth sign-in.

5. Playwright: best for browser automation

Microsoft's Playwright MCP lets an agent navigate websites, click, type, and read the page through the accessibility tree. It works on structure rather than screenshots, so it doesn't need a vision model.

It's the right pick for checking a signup flow on staging or walking through a web app the way a user would. One honest caveat comes from Microsoft itself: its README says coding agents increasingly prefer Playwright's CLI with skills because it uses fewer tokens. The MCP server still makes sense for exploratory automation, self-healing tests, and long-running agent workflows.

Best for agents that need to click through and test a web app the way a user would.

Docs: Playwright MCP

claude mcp add playwright npx @playwright/mcp@latest

6. Context7: best for current library documentation

Context7 pulls current documentation and code examples into the agent's context. Models are trained on docs that go stale, and Context7 fixes that for the library you're actually using.

It has two tools. resolve-library-id finds the right library, and query-docs fetches the docs. Mention the version in your request, say "Next.js 14", and it retrieves docs for that version. A free API key from the Context7 dashboard gets you higher rate limits. You still need to run the code it helps write, and coverage depends on what's in Context7's index.

Best for agents that need to write code against a specific library version without guessing at the API.

Docs: Context7

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

7. Supabase: best for database development

Supabase MCP lets an agent inspect a Supabase project and work with its database. It can list tables, run SQL, look at migrations, and generate TypeScript types. Other tool groups cover debugging, Edge Functions, branching, and docs. Storage tools are off by default.

The useful part is that the agent sees your real schema before it writes a query. The risky part is the same. Supabase's own guidance is to scope the server to one project with project_ref, turn on read_only=true, limit the tool groups with features, and keep production data out of it where you can.

Best for agents that need to understand a Postgres schema before they touch it.

Docs: Supabase MCP

claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp?features=docs%2Caccount%2Cdatabase%2Cdebugging%2Cdevelopment%2Cfunctions%2Cbranching"

8. Linear: best for issue and project tracking

Linear MCP lets an agent find, create, and update issues, projects, and comments. A support agent can read a bug report from an inbox and file it in Linear with repro steps and a link back to the conversation.

Linear hosts the server and supports OAuth 2.1 or an API key. There's a separate read-only endpoint at https://mcp.linear.app/mcp/readonly, which is a nice touch for reporting agents that should never touch a ticket.

Best for agents that need to turn requests into tracked issues and keep them current.

Docs: Linear MCP

claude mcp add --transport http linear-server https://mcp.linear.app/mcp

9. Ahrefs: best for SEO and AI search data

Ahrefs MCP gives an agent the same data you'd pull from the Ahrefs dashboard: backlinks and domain rating from Site Explorer, keyword volumes, SERP overviews, rank tracking, site audits, Google Search Console data, and Brand Radar for how often your brand shows up in AI answers.

That makes it the backbone for a marketing agent. It can find the keywords a competitor ranks for, check which of your pages lost traffic, or pull the pages AI assistants cite in your category, then hand that list to another tool.

It's a paid feature, available on Ahrefs plans from Lite up. Every call spends API units just like a direct API request, and workspace admins can cap monthly units per key. Watch that number if your agent loops.

Best for agents that need to answer SEO and AI visibility questions with real search data.

Docs: Ahrefs MCP

claude mcp add ahrefs https://api.ahrefs.com/mcp/mcp -t http

10. Stripe: best for billing and payment operations

Stripe MCP gives an agent access to your Stripe account through a handful of general tools that search, read, and write across the API, plus Stripe docs search and analytics. A support agent can look up a customer's subscription or a failed payment before drafting a reply.

Stripe makes a human confirm certain writes, including refunds and outbound payments. The agent gets a link, you approve it, and the approval expires after 24 hours. The bigger change is authentication. From October 31, 2026, Stripe MCP stops accepting full-access secret keys and restricted keys without the Agent tag, so reconnect with OAuth or create an agent API key. Test in a sandbox first.

Best for agents that need to look up payments, subscriptions, and invoices for support or ops work.

Docs: Stripe MCP

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

Stripe's recommended route is npm install -g @stripe/cli@latest followed by stripe agent setup, which configures the server and installs Stripe's skills for you.