Awesome MCP Servers: 27 Picks That Are Actually Good (2026)
The famous awesome-mcp-servers lists on GitHub index thousands of servers and recommend none of them. This is the opposite: 27 picks across 11 categories, each with a reason to choose it and an honest pointer to what to use instead. Curation signals: sustained install and reader activity on this directory, active maintenance, and distinctness. When you want the full index instead, our awesome MCP servers directory page lists every server by category with live counts.
On this page · 16 sections▾
TL;DR
If you only install three servers, make them Context7 (live library docs so your agent stops hallucinating APIs), GitHub (your code forge in the conversation), and Playwright (a real browser the agent can drive). Those three cover the workflows most developers actually run through an agent: write correct code, manage the repo, verify in a browser.
The rest of the list rounds out the map: scraping at scale (Bright Data, Firecrawl), free search (DuckDuckGo), databases (Supabase, PostgreSQL), project trackers (Jira, Notion, Obsidian), messaging (Slack, Telegram, Discord), local files and documents (Desktop Commander, Markitdown), agent memory (Mem0), infrastructure (Terraform, Coolify), and two picks that exist purely because they’re fun (PokePaste, FreeCAD). Everything below links to its canonical page on the directory, where the install configs live one click away.
One paragraph of orientation for anyone arriving cold: an MCP server is a small program that exposes tools — search the web, query a database, open a pull request — to any AI client that speaks the Model Context Protocol. Install one and Claude, Cursor, VS Code, and a dozen other clients all gain that capability at once. That write-once, use-everywhere property is why the ecosystem produced thousands of servers in two years, and why a curated list became necessary.
How we curated this
The awesome-list format has a known failure mode: it optimizes for inclusion, not judgment. A list with 3,000 entries answers “what exists?” and dodges “what should I install?” We already run the inclusive version — the awesome MCP servers page tracks thousands of servers across every category, sorted by GitHub stars and weekly installs, refreshed continuously. This post is the judgment layer on top of it.
Three rules governed every slot:
- Sustained traction, not spikes. Each pick shows months of steady install and reader activity on this directory, not a launch-week burst.
- Active maintenance. Abandoned servers break silently when clients update. Every pick here has a living upstream.
- Distinctness. Each server earns its slot by doing something nothing else on the list does. Where two servers overlap, we picked one and named the alternative in prose.
No placement is paid, and no pick is ours. License, transport, stars, and install configs live on each server’s card and canonical page — the prose below spends its words only on why.
Web search & scraping
The first capability most people bolt onto an agent: getting it information newer than its training data. Three picks, three cost shapes.
DuckDuckGo — free search, zero keys
The on-ramp. No API key, no account, no meter — install it and every client you own can search the web. Results are shallower than the commercial engines, but for “look this up before answering” workflows the marginal quality rarely justifies a paid key. Full setup and recipes in our DuckDuckGo MCP guide.
Bright Data — scraping that survives anti-bot walls
When the page fights back — CAPTCHAs, geo-blocks, JavaScript-rendered storefronts — Bright Data’s proxy network is the difference between data and a 403. It’s the heavyweight pick: metered, commercial, and worth it the moment your workflow touches sites that block datacenter IPs. Our Bright Data MCP guide covers the free Rapid tier and the cost traps.
Firecrawl — web pages as clean Markdown
Firecrawl’s pitch is LLM-ready output: it crawls a site and returns Markdown the model can reason over, instead of HTML soup that burns context. The right pick when the job is “ingest this docs site” rather than “search the web.” If you’re comparing crawlers, we ran the field side by side in Firecrawl vs AnyCrawl vs Crawlee vs Playwright.
Also strong: Brave Search for an independent index with a generous free tier, and Exa Search for semantic, meaning-based queries. The full field is in Best Web Search MCP Servers.
Docs & code context
The category that fixes the most common agent failure: confidently writing code against an API that changed two versions ago.
Context7 — live docs injected into every prompt
The most-read server on this entire directory, and the closest thing this list has to a default install. Add “use context7” to a prompt and the agent pulls current, version-specific documentation for the library you’re using — which kills the stale-API hallucination class outright. If you install one server from this post, install this one. Deep dive: Context7 MCP complete guide.
Serena — semantic code navigation for your own repo
Context7 knows the world’s libraries; Serena knows your codebase. It gives the agent language-server-grade tools — find symbol, find references, edit at the symbol level — so it navigates a large repo like an IDE instead of grepping and guessing. The practical effect: far fewer wrong-file edits in projects too big to fit in context. Setup and recipes: Serena MCP complete guide.
For understanding other people’s repos, DeepWiki turns any public GitHub project into queryable documentation — we compared the docs-context field in Context7 vs DeepWiki vs Ref vs Docfork.
Browser automation
A browser the agent can drive turns “I think this works” into “I watched it work.” Three picks for three browser philosophies.
Playwright — the default browser for agents
Microsoft’s accessibility-snapshot model is the reason this is the category default: the agent reads the page as structured text rather than screenshots, which is faster, cheaper, and survives layout changes. Cross-browser coverage means your test isn’t silently Chromium-only. Start with our Playwright browser MCP guide.
Chrome DevTools — performance numbers nothing else has
Not a Playwright substitute — a complement. It attaches to Chrome’s debugging protocol, so the agent can run performance traces and read Core Web Vitals, console errors, and network waterfalls. When the question is “why is this page slow?” this is the only pick in the category that can answer with data. Head-to-head: Chrome DevTools MCP vs Playwright MCP.
Selenium — meet your existing test estate where it is
If your team has a decade of Selenium grid infrastructure and WebDriver suites, an agent that speaks Selenium slots into that estate instead of demanding a migration. That’s the whole pitch, and for a lot of enterprise teams it’s decisive. New projects should start with Playwright; existing Selenium shops should read the Selenium MCP complete guide. (Puppeteer remains the minimalist Chromium option.)
Databases
Letting an agent query your data is the highest-leverage, highest-risk integration on this list. Both picks here have sane answers to the risk half.
Supabase — your whole backend, one server
The official Supabase server covers the database and the platform around it — tables, migrations, edge functions, logs — so the agent can reason about your backend as a system rather than a connection string. Read-only mode and project scoping are first-class, which is exactly the guardrail shape you want before pointing a model at production. In practice the winning prompts are the analytical ones — “which queries against this table got slower this week, and is an index missing?” — where the agent joins schema knowledge to logs in a way no dashboard does.
PostgreSQL — talk to any Postgres, anywhere
The general-purpose pick: schema inspection, queries, and table management against any Postgres you can reach — RDS, Neon, a docker-compose service, the legacy box under someone’s desk. Less platform awareness than the Supabase server, more universality. If your data lives in documents instead, MongoDB has the equivalent server.
Dev & git
The category where agents stop being chatbots and start being coworkers: repos, reviews, containers.
GitHub — the indispensable one
Issues, pull requests, CI runs, code search — the official server puts the entire GitHub surface in the conversation, and “triage my open issues and draft replies” stops being a context-switch carnival. If your team lives on GitHub, this is the second server you install after Context7, and it’s not close. The pattern that sells it: a code-review pass where the agent reads the diff, checks the linked issue, scans the failing CI log, and drafts review comments — four GitHub tabs collapsed into one conversation.
Docker — containers in plain language
Compose stacks, logs, container lifecycle — managed conversationally. The underrated use is debugging: “why is the api container restarting?” becomes one prompt that reads logs, inspects state, and proposes the fix, instead of four terminal tabs.
Bitbucket — for the Atlassian half of the industry
GitHub-centric lists pretend Bitbucket shops don’t exist; they’re a large fraction of enterprise teams. This server covers repositories, pull requests, and pipelines over the REST API, and pairs naturally with the Jira pick below. Setup: Bitbucket MCP complete guide. For error tracking alongside either forge, Sentry is the standard add-on.
Productivity & project management
Where the “agent as coworker” framing gets tested against real backlogs and real notes.
Notion — the workspace API, agent-shaped
If Notion is your team’s source of truth, this server makes it readable and writable from any MCP client: meeting notes become database rows, research dumps become structured pages, and the agent files things where they belong instead of into a chat transcript you’ll never reopen.
Jira — backlog hygiene without the dashboard
The unglamorous pick that saves the most minutes per week. Creating tickets from a bug description, updating statuses after a deploy, summarizing a sprint — all chores the agent does faster than the Jira UI loads. Recipes and auth setup: Jira MCP complete guide. If your tracker is Linear, the comparison in Jira vs Linear vs Trello vs Productboard MCP settles it.
Obsidian — your second brain, queryable
Years of accumulated notes become something the agent can actually use: “what did I conclude about this last time?” is now answerable. Because vaults are local Markdown, your knowledge base stays on your disk — the rare productivity integration with no cloud dependency. Full walkthrough: Obsidian MCP complete guide.
Messaging
Three platforms, three different relationships between agent and conversation.
Slack — where work conversations go to be forgotten
The killer prompt is retrieval: “find the thread where we decided on the retry policy” works across channels you’d never scroll manually. Posting summaries and standup updates is the bonus. For teams that make decisions in Slack — most teams — this is the institutional-memory server.
Telegram — the personal automation channel
Telegram’s open bot API makes it the easiest place to give yourself an agent-powered inbox: monitor channels, digest group chats, send yourself alerts from any workflow. Where Slack is for work, this is the pick for personal pipelines. Setup: Telegram MCP complete guide.
Discord — community management at scale
Built for the server-admin job: messaging, moderation, roles, and events across a surface of 46+ tools. If you run a community — open-source project, game server, paid group — this turns recurring moderation chores into prompts. Guide: Discord MCP complete guide.
Files & documents
The local layer: what the agent can touch on your machine, and how documents become text it can reason over.
Desktop Commander — your machine, under agent control
Terminal commands, file search and editing, and process management in one server — the closest thing to giving the agent hands on your workstation, usable from clients that don’t ship their own shell access. Power and risk scale together here; read the guardrails section of our Desktop Commander MCP guide before granting it a wide directory scope.
Markitdown — every document becomes Markdown
Microsoft’s converter handles the unglamorous step every document workflow starts with: PDF, Word, PowerPoint, Excel in; clean Markdown out. As an MCP server it means “summarize this quarterly report” works on the actual file, not a copy-paste of it. Details: Markitdown MCP complete guide. For plain sandboxed file access, the reference Filesystem server is the minimal option.
Memory & reasoning
Two servers that upgrade the agent itself rather than connecting it to something.
Mem0 — the agent that remembers you
Persistent, semantically searchable memory across sessions: preferences, decisions, project context that survives the conversation ending. It’s the difference between an agent you brief every morning and one that already knows. The memory-layer field is crowded; we compared it in Mem0 vs Letta vs Zep vs Cognee.
Sequential Thinking — structured reasoning on demand
A reference server that gives the model a scratchpad protocol: decompose the problem, revise earlier steps, branch when needed. On gnarly multi-step problems — migrations, debugging, anything with interacting constraints — the visible improvement in answer quality is why this stays one of the most-installed servers anywhere. It costs nothing, runs locally, and has no failure mode worse than an unused tool, which makes it the rare “install it everywhere” recommendation we can give without caveats.
Deploy & infrastructure
Infrastructure rewards the agent that reads documentation perfectly and punishes the one that guesses.
HashiCorp Terraform — IaC without the doc-tab sprawl
The official server gives the agent real-time access to provider docs and module specs from the Terraform Registry — so generated HCL references arguments that exist in the provider version you’re pinned to. One of the most-viewed servers on this directory, because the alternative is an agent inventing resource attributes.
Coolify — self-hosted PaaS, conversationally
Coolify is the self-hoster’s Heroku replacement; its MCP server means deployments, databases, and service status on your own VPS are one prompt away. The pick for people who left the big clouds on purpose. Setup: Coolify MCP complete guide. On the managed side, Cloudflare’s server covers Workers and the rest of its edge platform.
Fun & niche
Every honest awesome-list keeps a shelf for servers that exist because someone wanted them to.
PokePaste — competitive Pokémon teams from a prompt
Draft VGC teams with an agent and ship them as pokepast.es links. Silly? It’s also one of the cleanest demonstrations of what MCP is for: a niche community had a data format, someone wrapped it in tools, and now every AI client speaks it. The write-up: PokePaste MCP complete guide.
FreeCAD — parametric CAD by conversation
Drive real CAD modeling — sketches, solids, scripted edits — through natural language via FreeCAD’s Python scripting layer. The most surprising entry on the list to watch in action, and a glimpse of agents operating professional desktop software generally. Guide: FreeCAD MCP complete guide.
How to choose: build a stack, not a hoard
The mistake this list invites is installing all 27. Don’t. Every connected server adds its tool descriptions to the model’s context window, and past a handful the agent gets measurably worse at picking the right tool — we wrote up the failure mode and fixes in the MCP context bloat post. Build per-job stacks instead:
- Writing code: Context7 + GitHub + your database server. Add Serena when the repo outgrows the context window.
- Testing and QA: Playwright + GitHub. Add Chrome DevTools when performance questions show up.
- Research and content: DuckDuckGo (or Brave) + Firecrawl + Notion or Obsidian to file the results.
- Ops and infra: Terraform + Docker + the platform you deploy to (Coolify, Cloudflare).
- Team workflows: Slack + Jira (or Linear) + GitHub or Bitbucket — matching whichever half of the Atlassian divide you live on.
A second heuristic when two servers overlap: prefer the official vendor server when one exists, because it tracks API changes the day they ship; prefer the community server when it’s meaningfully better designed (Serena and Desktop Commander both beat thinner official equivalents in their niches). And before adding anything, ask whether your client already does the job natively — Claude Code ships its own file and shell tools, so Desktop Commander matters most in clients that don’t.
Two cross-cutting picks upgrade any stack: Mem0 if you run long-lived projects through the same agent, and Sequential Thinking if your problems are multi-step. And if budget is the constraint, nearly every category above has a free floor — mapped in Best Free MCP Servers.
The 10-minute starter stack
Context7 for correct code, GitHub for your repos, DuckDuckGo for free search, Playwright when you need a browser. Four servers, no API keys except a GitHub token, and you’ve covered most of what people actually use MCP for. Expand from the categories above only when a real job demands it.
Frequently asked questions
What is the awesome-mcp-servers list?
awesome-mcp-servers started as a community-maintained GitHub awesome-list — a long, categorized index of Model Context Protocol servers, the most-starred being punkpeye/awesome-mcp-servers. This post is the editorial version of that idea: instead of indexing everything, it picks 27 servers across 11 categories and explains why each one earns the slot. For the full searchable index, see our awesome MCP servers directory page.
What are the best MCP servers in 2026?
By install activity and reader traffic on this directory, the consistent leaders are Context7 (live library docs), the official GitHub MCP server, Playwright for browser automation, Bright Data for scraping at scale, Desktop Commander for local file and terminal work, and Mem0 for persistent agent memory. 'Best' depends entirely on the job — this post groups picks by category so you can match a server to the work in front of you.
How many MCP servers should I install at once?
Two to four, picked by job. Every connected server adds its tool descriptions to the model's context, so wiring up fifteen servers degrades reasoning and burns tokens before you type a word. A sane default stack: one search or docs server, one server for your code forge, and one for the system you query most (database, project tracker, or notes). Add and remove per project rather than accumulating.
Are the MCP servers on this list free?
Most are. The open-source servers (Playwright, DuckDuckGo, Desktop Commander, Sequential Thinking, Serena, Markitdown, and others) cost nothing. Servers fronting commercial APIs — Bright Data, Firecrawl, Datadog-class tools — are free to install but metered by the underlying service, usually with a free tier. We flag the cost shape per pick; for a zero-budget stack, see our best free MCP servers roundup.
Which MCP server should I install first?
If you write code: Context7, because outdated API knowledge is the failure you hit on day one, and live docs injection fixes it for every framework at once. If your work is research: DuckDuckGo, because it gives any client free web search with no API key. Both take under two minutes to wire into Claude, Cursor, or VS Code.
How is this list different from the GitHub awesome-mcp-servers repo?
The GitHub repos are inclusive indexes — thousands of entries, community-submitted, minimal commentary. Useful for discovery, weak for decisions. This list inverts that: 27 entries, each with an opinionated reason to pick it and an honest pointer to alternatives. When you want breadth, our directory page lists every server by category with live install counts and GitHub stars.
How were these 27 servers chosen?
Three signals: sustained reader and install activity on this directory (not a one-week spike), active maintenance of the underlying server, and distinctness — each pick has to do something the others on the list don't. No placement is paid. Official vendor servers were preferred where they exist; community servers made the list where they're flatly better or the only option.
Sources
- /awesome-mcp-servers — the full, continuously updated index this post curates from, with live stars and install counts
- /servers — browse and filter the complete catalog; every pick above links to its canonical page there
- github.com/punkpeye/awesome-mcp-servers — the community awesome-list that defined the format this post responds to
- github.com/modelcontextprotocol/servers — the official reference servers (Filesystem, Sequential Thinking, and the archived originals)
- MCP.Directory install and view telemetry, June 2026 — the traction signal behind every pick; described in shape, not raw numbers, because counts move daily
Deep dives for featured servers