11 Best MCP Servers in 2026, by Category

"Which MCP servers should I actually install?" is the question everyone asks the week after they learn what an MCP server is. The honest answer: fewer than the directories suggest. Registries now list thousands of servers, but a handful cover the workflows most people run daily — and connecting too many at once degrades tool selection, because every server's tool descriptions compete for the model's attention.

This list is the short version: the best MCP servers by category as of 2026, what each one is genuinely good at, and the limits worth knowing before you connect it. Every entry is either an official first-party server or the clear community standard.

How we picked: first-party maintenance (or de facto community standard), meaningful capability through typed tools rather than a thin demo, active upkeep through 2026, and a real security model — scoped credentials, not account-wide keys.

The eleven: GitHub, Playwright, Chrome DevTools, and Context7 (coding); AFFiNE, Notion, and Obsidian (knowledge); Slack and Sentry (communication and ops); Supabase and n8n (data and automation).

Coding and development

GitHub

The reference example of a first-party server done well. Repos, issues, pull requests, and code search become tools an agent can call, so "summarize the open bugs tagged regression and draft a fix plan" runs against live project state instead of a stale paste. It is available as GitHub's hosted remote server or as a local run, and fine-grained token support means you can scope it to a single repo.

Limits: tool count is large; if your client struggles with tool selection, enable only the toolsets you use.

Playwright

Microsoft's browser-automation server gives agents a real browser: navigate, click, fill forms, run accessibility snapshots. It is the backbone of "test this flow end-to-end" and "why does checkout break on mobile viewports" conversations. Because it drives an actual browser engine, results reflect what users see, not what the HTML implies.

Limits: long sessions can accumulate state; start fresh for reproducible runs.

Chrome DevTools

Where Playwright drives the page, the Chrome DevTools server inspects it: performance traces, network requests, console errors, DOM state. For frontend debugging with an agent — "what's blocking LCP on this page" — it is the difference between guessing and reading the profiler.

Limits: most useful when you can reproduce the issue locally; it inspects a browser you run.

Context7

Solves the quiet killer of AI coding: outdated API knowledge. Context7 serves versioned, current documentation for thousands of libraries, so generated code calls the API that exists today rather than the one from the model's training data.

Limits: coverage is deepest for popular open-source libraries; niche internal SDKs still need your own docs — which is what a knowledge-base server is for.

Knowledge and notes

AFFiNE

AFFiNE's built-in MCP server connects assistants to the place your team already writes: docs and whiteboards in one workspace. The distinguishing capability is search — keyword and semantic — that also indexes whiteboard canvases, so an assistant can find the architecture sketch or sticky-note cluster where a decision actually happened, not just the meeting note that mentions it.

It is also first-party infrastructure rather than a bolt-on. Credentials are created in workspace settings, scoped to one workspace with your permissions, read-only by default, with per-client revocation and expiry. Setup is copy-paste (the app generates the full JSON configuration), and there is no separate charge on AFFiNE Cloud or self-hosted instances with AI features enabled.

Limits: document reads return page content as Markdown; whiteboard content is searchable rather than readable as a canvas. Write access is rolling out progressively rather than on by default.

Notion

The official hosted server exposes pages and databases to assistants, with OAuth-scoped access. If your team's structured content lives in Notion databases, it is the straightforward choice.

Limits: the current official server is hosted-only — there is no supported self-hosted path (an earlier open-source server is deprecated) — and canvas-style content is not part of the model.

Obsidian (community)

Community servers built on Obsidian's REST API plugin let agents search and read a local vault — attractive if your notes are plain Markdown files on disk.

Limits: the desktop app must be running for the API to respond, and you are trusting a community integration with your vault; read the permissions before connecting. Full setup and trade-offs: Obsidian MCP guide.

Communication and operations

Slack

Channels, threads, and search as tools. The practical unlock is institutional memory: "what did we decide about the pricing page in March" stops being an archaeology project. Scope the token to the channels the assistant actually needs.

Limits: search quality inherits Slack's own search; very large workspaces benefit from channel-scoped queries.

Sentry

Error tracking meets agents: query issues, read stack traces, correlate releases. Pairing it with GitHub's server closes the loop from "alert fired" to "here is the diff that likely caused it."

Limits: deep triage still wants a human confirming the fix; treat agent conclusions as leads.

Data and automation

Supabase

Database schema, queries, and edge functions as typed tools. For prototyping with an agent — "add a table for invoices and wire up the query" — it removes the copy-paste loop between chat and SQL editor.

Limits: give it a development branch, not production credentials. Always.

n8n

One connection that fans out to hundreds of downstream integrations through workflows you control. Instead of installing ten single-purpose servers, expose three or four curated workflows as tools.

Limits: you are maintaining the workflows; the server is only as good as what you wire into it.

Quick comparison

How to choose (and how many to connect)

• Start from repeated pastes. Whatever you copy into AI chats weekly is your first server. For most knowledge workers that is their notes; for most developers, their repo.
• Three to five servers beats fifteen. Every connected server adds tool descriptions to the model's context. Curate per project: a coding profile (GitHub + Playwright + Context7) and a knowledge profile (AFFiNE + Slack) cover a remarkable amount of work.
• Prefer first-party. A server maintained by the product it connects to inherits real auth and gets updates when the product changes. For anything community-built, read what it can access before connecting it to sensitive data.
• Check the registry before building. The official MCP registry and directories like PulseMCP and Glama list thousands of servers; the connector you are about to write probably exists.

FAQ

What are the best MCP servers to start with?

Start with one server for your code and one for your knowledge. Developers typically connect GitHub plus a docs server like Context7; knowledge workers connect their notes first — AFFiNE's built-in server makes docs and whiteboards searchable by any MCP client in a few minutes, with no separate MCP charge when AI features are enabled.

Are there free MCP servers?

Most official servers are free to connect and use with an existing account. AFFiNE's MCP server has no separate charge on Cloud or self-hosted instances when AI features are enabled, GitHub's works with a free account, and open-source servers like Playwright cost nothing to run locally. Costs, where they exist, come from the underlying product's plan.

How many MCP servers should I connect at once?

Three to five per profile works best in practice. Each server adds tool descriptions the model must weigh on every request, so twenty connected servers make tool selection slower and less accurate. Curate a small set per workflow and disable what you are not using.

Is there an official list of MCP servers?

Yes — the official MCP registry launched in preview in late 2025, and the modelcontextprotocol GitHub organization maintains reference servers. Community directories such as PulseMCP, Glama, and mcp.so track the wider ecosystem, which now spans thousands of servers; how the upstream catalog works is covered in our MCP Registry guide.

Which MCP server works for team knowledge bases?

AFFiNE's built-in MCP server is designed for exactly that: workspace-scoped credentials created in settings, keyword plus semantic search across documents and whiteboards, and read-only access by default. Notion's official hosted server is the alternative when your team's content lives in Notion databases.

---

The pattern across every good server on this list: it connects an assistant to a system of record you already maintain. If your team's decisions live in docs and on whiteboards, make them searchable first — it is the highest-leverage connection most teams can make, and it takes about five minutes.