Most "best MCP servers" lists mix a handful of vendor-maintained servers in with third-party wrappers of varying quality, ranked by not much more than how often they come up in search. That's not a useful comparison once you're deciding what to actually connect to an agent that can read your code, your database or your production error logs.
This list keeps to one rule: every entry is maintained by the company whose name is on it. GitHub's server comes from GitHub. Sentry's comes from Sentry. That's not a guarantee of quality by itself, but it means the tool surface, the auth model and the security posture are the vendor's own responsibility, not a third party's best guess at their API from the outside.
How these were chosen
Eight servers, each verified directly against its own repository or documentation rather than secondary write-ups. Every one is either the officially maintained server for its platform or, in Playwright's and Context7's cases, the closest thing to an official reference implementation for that category. Excluded: community wrappers around vendor APIs the vendor itself doesn't maintain, servers without a public repository to actually read before connecting them, and anything whose tool list couldn't be independently confirmed.
If you're new to MCP generally, the short version is that a server exposes tools an agent can call directly, database queries, API calls, browser actions, rather than the agent working from what you paste into chat or guessing at an unfamiliar API from training data.
The best MCP servers for developers in 2026
GitHub MCP Server
Best overallGitHub's own MCP server for repos, issues, pull requests, Actions and code scanning.
- Best for
- Any coding agent that needs to read or act on a GitHub repository
- Pricing
- Free and open source; uses your existing GitHub OAuth or personal access token permissions
- Organised into toolsets (repos, issues, pull requests, Actions, code security) so you can enable only what you need
- Hosted remote server at api.githubcopilot.com needs no local process
- One-click install buttons for VS Code and Claude Desktop
- Full toolset list is broad enough that scoping it deliberately matters for a shared or lower-trust environment
- Local mode requires Docker or a Go build if you don't want the hosted version
The GitHub MCP server is the obvious starting point for almost any coding agent setup, and the toolset design is why it earns the top spot rather than just default familiarity. Instead of one undifferentiated bundle of tools, it's organised into toolsets, repos, issues, pull requests, Actions, code security, discussions, projects, notifications, so you can enable exactly what a given agent needs and nothing it doesn't. The hosted remote server at api.githubcopilot.com/mcp/ means most people never need to run anything locally at all.
The honest limitation: the full toolset is broad, and broad access is exactly what you don't want handed to a low-trust agent or a shared environment by default. Use --toolsets to scope it down rather than accepting everything.
Playwright MCP
Best for browser automationMicrosoft's MCP server for driving a real browser through Playwright, using accessibility snapshots instead of screenshots.
- Best for
- Agents that need to test, scrape or automate a web app end to end
- Pricing
- Free and open source; runs locally, no account required
- Structured accessibility snapshots instead of screenshots, cheaper and more reliable for an agent to parse than pixels
- Covers navigation, forms, tabs, network mocking, cookies and PDF generation in one server
- One npx command to run, no separate browser install step beyond Node 18+
- Test assertions and some DevTools features are opt-in extras, not on by default
- Purely local, no hosted remote option if you want to avoid running it yourself
Playwright MCP's real advantage over screenshot-based browser automation is the accessibility snapshot approach: an agent gets a structured representation of what's actually on the page, which is cheaper to process and far more reliable to act on than asking a model to interpret pixels. Between navigation, form filling, network mocking, cookie and storage management, and PDF generation, it covers most of what a browser-testing or scraping agent needs from a single server.
It's local-only, though, with no hosted remote option documented. If you want browser automation without running anything yourself, this isn't that; it needs Node 18+ and a real Playwright browser install on the machine running it.
Context7
Best for current library docsInjects up-to-date, version-specific documentation and code examples into an agent's context, instead of relying on training data that's already stale.
- Best for
- Agents writing code against fast-moving libraries and frameworks
- Pricing
- Requires a Context7 API key, generated free during setup; the repository doesn't document a paid tier
- Directly addresses a real, common failure mode: an agent confidently citing an API that changed or never existed
- Two simple tools (resolve-library-id, query-docs) keep the surface area small
- One command, npx ctx7 setup, handles OAuth and installs the matching skill alongside it
- Only as good as the documentation Context7 has indexed for a given library
- Hosted-only; there's no offline or self-hosted mode documented in the repo
Context7 solves a specific, common problem: a model's training data goes stale the moment a library ships a new major version, and the model has no way to know that from inside a conversation. Context7 fixes this by injecting current, version-specific documentation directly into context via two tools, resolve-library-id and query-docs, rather than leaving the agent to work from what it already "knows."
It's hosted-only, so you're dependent on Context7's own index having decent coverage of whatever library you're working against, and there's no offline mode if that index doesn't cover something obscure.
Supabase MCP
Best for a Postgres-backed projectSupabase's own MCP server for querying data, managing tables and fetching project config from an agent.
- Best for
- Teams already on Supabase who want an agent to work directly against their database
- Pricing
- Free server; billed through your existing Supabase project plan
- Hosted endpoint (mcp.supabase.com) with OAuth 2.1, no local process needed for a hosted project
- Local dev mode against the Supabase CLI keeps agent access out of production entirely while testing
- npm package available for direct integration into custom agent stacks, including the Vercel AI SDK
- Connecting an LLM directly to a database is real risk if credentials or scopes aren't set deliberately; Supabase's own docs say as much
- The repository points you to Supabase's separate MCP docs page for the full tool list rather than documenting it inline
Supabase's server is the straightforward pick if your backend already runs on Supabase: query data, manage tables and read project config, all through either a hosted OAuth endpoint or, for anyone nervous about pointing an agent at production, a local mode that runs against the Supabase CLI instead. That local/hosted split is a genuinely useful safety valve most database-adjacent MCP servers don't offer as cleanly.
Supabase's own documentation is upfront that connecting an LLM to your database carries real risk, and it's worth taking that warning at face value rather than as boilerplate. Scope credentials narrowly and read the security guidance before pointing this at anything with real data in it.
Sentry MCP
Best for debugging from production errorsSentry's own MCP server, letting an agent search issues, traces and performance data with natural-language queries.
- Best for
- Debugging workflows where the agent needs to see what's actually failing in production
- Pricing
- Free server; requires a Sentry account and access token, billed on your existing Sentry plan
- AI-powered search converts natural language into Sentry's own query syntax, rather than making the agent learn it
- Available as a Claude Code plugin for one-step install, alongside plain stdio and a hosted remote option
- Supports self-hosted Sentry instances, not just the SaaS product
- The AI-powered search tools need their own LLM provider configured (OpenAI, Anthropic, Azure or OpenRouter), separate from whichever model is driving your agent
- Explicitly designed around human-in-the-loop coding workflows rather than general-purpose Sentry API coverage
Sentry's server turns "check what's actually broken in production" into something an agent can do directly, converting natural-language queries into Sentry's own query syntax rather than requiring the agent to already know it. It ships three ways: a Claude Code plugin for the fastest setup, a local stdio server, or the hosted mcp.sentry.dev endpoint, and it supports self-hosted Sentry instances too, not just the SaaS product.
One wrinkle: the AI-powered search tools need their own LLM provider configured separately from whichever model is running your agent, which is an extra piece of setup most of the other servers on this list don't require.
Stripe MCP
Best for agents that touch billingStripe's hosted MCP server for building agents that can read or act on payment data through OAuth.
- Best for
- Teams building support or billing-ops agents against a Stripe account
- Pricing
- Free server; billed through your existing Stripe account and its usual transaction fees
- Hosted directly by Stripe at mcp.stripe.com, no separate infrastructure to run
- OAuth-based access keeps the connection scoped to your account rather than a shared API key
- Sits alongside Stripe's Agent Toolkit for teams building custom agent integrations beyond MCP alone
- The tool list itself lives on Stripe's own docs site rather than in a public repo you can read end to end before connecting it
- Anything touching live payment data deserves the most conservative credential scoping on this whole list
Stripe hosts its MCP server directly, no local install to manage, authenticated over OAuth rather than a shared API key sitting in a config file. That's a sensible default for anything touching live payment data, and it sits alongside Stripe's separate Agent Toolkit for teams that want to build more custom agent integrations than MCP alone covers.
The trade-off is transparency: the exact tool list lives on Stripe's own documentation rather than in a public repository you can read start to finish before connecting it, which matters more here than almost anywhere else on this list given what's at stake if scoping goes wrong.
Notion MCP
Best for docs and knowledge basesNotion's official server for reading and writing pages, databases and comments from an agent.
- Best for
- Teams that keep specs, runbooks or knowledge bases in Notion and want an agent to work with them directly
- Pricing
- Free server; requires a Notion internal integration token, scoped per page or database
- 22 documented tools covering data sources, page markdown, search and comments
- Works over plain npx or a published Docker image (mcp/notion)
- Fine-grained access: you grant the integration token specific pages and databases, not your whole workspace by default
- Notion's own README says it may sunset this local server in favour of its hosted Notion MCP, and that issues here aren't actively monitored
- Worth starting with the hosted Notion MCP (OAuth-based) instead of this repo unless you specifically need the local, token-based flow
Notion's server is a solid pick if your team's specs, runbooks or internal docs already live in Notion. Twenty-two documented tools cover data sources, page content as markdown, search and comments, with access scoped per integration token to specific pages or databases rather than your whole workspace.
Read Notion's own README carefully before betting long-term infrastructure on this exact repo: Notion is explicit that it's prioritising its newer, hosted Notion MCP (OAuth-based) and "may sunset this local MCP server repository in the future," with issues and pull requests here no longer actively monitored. For a new setup, start with the hosted version unless you specifically need the local, token-based flow this repo provides.
Cloudflare MCP servers
Best if your infrastructure is on CloudflareNot one server but 16+ domain-specific ones, covering Workers, DNS analytics, security auditing and more.
- Best for
- Teams running Workers, DNS or edge infrastructure on Cloudflare who want an agent to inspect or manage it
- Pricing
- Free servers; billed through your existing Cloudflare account and API credentials
- Split by domain (Workers, analytics, security, content, infrastructure) so you connect only what a given agent actually needs
- Hosted remote endpoints (docs.mcp.cloudflare.com and others) need no local process
- Covers unusually specific ground for a vendor server: audit logs, CASB scanning, browser rendering, AutoRAG
- 16+ separate servers is more to evaluate and choose between than a single unified one
- Each domain-specific server needs its own Cloudflare API credentials configured
Cloudflare didn't ship one server, it shipped sixteen-plus, split by domain: Workers development, DNS and traffic analytics, security auditing, browser rendering, and more. If your infrastructure actually lives on Cloudflare, that granularity is a feature, connect only the specific servers an agent needs rather than one server with everything bundled in.
The trade-off is exactly that granularity: evaluating and configuring credentials for the right subset out of sixteen-plus options is more upfront work than picking a single unified server, and it's worth reading Cloudflare's own repository to work out which of them actually apply to your stack before connecting any.
Local vs hosted: which to pick
Six of the eight servers above now offer a hosted remote endpoint, and it's worth defaulting to it rather than running a local process out of habit. A hosted server means nothing to keep patched, nothing that stops working because your laptop went to sleep mid-task, and usually OAuth-based auth instead of a long-lived API key sitting in a config file on disk. Cloudflare, GitHub, Context7, Supabase, Sentry and Stripe all fit that pattern now.
Local, stdio-run servers still make sense in three situations: when there's no hosted option at all, which is currently true of Playwright's browser automation; when you specifically want to keep an agent's access to a lower-trust environment like a dev database rather than production, which Supabase's local CLI mode is built for; or when a vendor's hosted offering is genuinely newer and less proven than its local one, which is closer to Notion's situation than the reverse. Read the specific trade-off for whichever server you're picking rather than assuming one mode is always right.
Servers worth knowing about, not yet on this list
A handful of other vendor-maintained servers come up constantly in MCP discussions, Linear's, Figma's and Vercel's among them, but this list only includes servers verified directly against a repository or documentation page reachable at the time of writing. Where that verification wasn't possible this run, the honest move is to leave the server off rather than describe it from secondary summaries. If you rely on one of those platforms, check for its own official MCP server directly on the vendor's site or GitHub organisation; the installation pattern will look like the ones above regardless of which vendor it's from.
How to actually connect one
The exact steps vary by client, but the shape is the same everywhere: point your agent at either a hosted URL or a local command, with credentials scoped as narrowly as the server allows. In Claude Code specifically:
# Remote server (most of the list above)
claude mcp add --transport http <name> <url>
# Real example: connect the hosted GitHub server
claude mcp add --transport http github https://api.githubcopilot.com/mcp/
# Local stdio server (Playwright, or Notion's local mode)
claude mcp add --env NOTION_TOKEN=your-token --transport stdio notion \
-- npx -y @notionhq/notion-mcp-server
The same servers work the same way in any MCP-compatible client, since MCP is the shared standard underneath. See the platforms page for which agents support MCP natively.
What to check before you connect one
Confirm it's actually the vendor's own server, not a third-party project with a similar name. Every server on this list is published under the vendor's own GitHub organisation or domain; that's a meaningful signal a random fork or clone doesn't carry.
Scope credentials to the minimum the task needs. A read-only token beats a full-access one whenever the server supports the distinction, and several on this list, GitHub and Notion especially, let you scope access down to specific repos or pages rather than an entire account.
Read the repository before connecting anything that touches money, production data, or credentials. The same instinct this site applies to agent skills applies here: an MCP server is code that runs with real permissions, and the fact that a vendor publishes it doesn't remove the value of a quick read before you trust it with your Stripe account or your production database.
Prefer the hosted endpoint when one exists and you don't have a specific reason to self-host. It's one less process to keep running and patched, and OAuth-based hosted auth is generally a tighter model than a long-lived API key sitting in a local config file.
Where to go next
For how MCP servers and agent skills divide responsibility, an MCP server exposes capability, a skill tells an agent how to use it well, see Agent Skills vs MCP. For the newest MCP specification and what it changes about how servers like these actually talk to a client, see What changed in the MCP 2026-07-28 spec. Browse platforms with native MCP support at getclaudeskills.com/platforms, or skills built to work alongside a specific MCP server at getclaudeskills.com/skills.
Verified 21 August 2026 directly against each server's own GitHub repository, plus Stripe's own MCP documentation page for the one entry without a public repo to read. Tool lists and installation steps reflect what each vendor documents as of that date; check the vendor's own page before connecting, since MCP servers update independently of this article.
