Best MCP servers, with maintained and archived status
The MCP servers with the best evidence and upkeep as of 22 October 2026 are Playwright for browser control, Context7 for current library docs, n8n-MCP for building n8n workflows, and Figma's official server for design-to-code. Before you install a server from an older list, check it is not one of the reference servers the MCP project archived in May 2025, such as the old Slack, GitHub, Puppeteer or Brave Search servers [3].
Last verified 2026-10-22 · GitHub status checked through the GitHub API on that date
An MCP server gives an AI agent a set of tools for one outside system, such as a browser, a design file or an ad account, over the Model Context Protocol [1]. Claude Code adds one with claude mcp add[2]. We took every MCP server named in the 43 saved posts in the DejaViewed corpus that mention MCP (9 S-tier, 19 A-tier, 7 B-tier, 8 C-tier). We checked each server's repo or official docs page and gave it a verdict: 4 keep, 18 try, 3 skip.
Status comes from the GitHub API on 22 October 2026 [5]. Maintained means a commit in the last 90 days, slow updates means 90 to 365 days, and archived means the owner archived the repo. Official, hosted means the vendor of the app runs the server and documents it, with no public repo to check. Install commands are copied from the entry, the README or the vendor's docs.
A server runs with whatever access you give it. Servers that can spend money, such as ad accounts, post publicly, or scrape sites that forbid it carry the most risk, and the reasons below say so where it applies.
Screens and replays TradingView charts for backtests from the agent.
Playwright MCP
KeepMaintainedcommunity
Drives a real browser by natural language: navigate, click, fill forms, screenshot pages so the agent can test or operate websites.
Why keep: Most-cited server in the corpus (7 entries); Microsoft-maintained, committed today. DXGRONttjhp argues the Playwright CLI is leaner for scale. Agent acts in logged-in browser sessions.
claude mcp add playwright npx @playwright/mcp@latest
Figma MCP (Dev Mode MCP server)
KeepOfficial, hostedofficial
Lets the agent read Figma frames, variables and components so generated code matches the design, and push work back.
Why keep: Four entries use it in design-to-code loops (two pair it with Playwright). Official Figma server, hosted; DV09StsDI2d notes it burns Pro usage quickly.
Gives the agent knowledge of n8n's nodes so it can design, validate and deploy n8n automation workflows.
Why keep: Named with repo in DUzg0n1jRK-; DX62rNWxOB2 uses it to build a Gmail workflow. Community-built, very active. With an n8n API key it can create/edit live workflows.
Searches and generates shadcn/Tailwind React components from the 21st.dev library inside the coding agent.
Why try: Two entries recommend it to escape generic AI UI. Official 21st.dev server; Magic MCP renamed to 21st MCP, old npx package proxies. Needs API key.
Lets an LLM inspect and build Blender scenes, generate assets (e.g. via Tripo) and arrange them from a reference image.
Why try: Two entries demo full scene assembly. Repo renamed ahujasid/blender-mcp to ahujasid/mcp-for-blender; old uvx name still works. Community, not Blender Foundation; runs arbitrary Python inside Blender.
Exposes Stitch-generated UI screens to the coding agent so it can build production code from those mockups.
Why try: Two entries export Stitch designs to Claude Code via its MCP setup button. Official Google hosted server; davideast/stitch-mcp is a separate community CLI.
Reads Meta ad performance and creates, edits and launches campaigns and creatives from the agent.
Why try: Two entries wire it in via custom connector. Meta's official hosted server (mcp.facebook.com/ads). Write access to live ad accounts: agent can spend money; start read-only.
Runs Apify Store scrapers so the agent can pull data on videos, channels and posts across social platforms.
Why try: DVmw74XEbk7 uses it as a social-media analyst. Official Apify server, active; hosted at mcp.apify.com. Pay-per-run scraping; check each platform's terms.
Reads, triages and drafts replies to email from inside Claude.
Why try: DZVXehQRe3T auto-answers 13 of 16 support emails with it plus FAQ markdown. Anthropic's built-in connector, not Google's. Full inbox access; incoming email is untrusted input.
Converts PDFs, images, audio and web pages into Markdown the agent can read.
Why try: One shortlist entry; last commit May 2026. Microsoft's markitdown (very active) covers similar conversion if this stalls. Local file access only.
Persistent long-term memory for agents, exposed as an MCP server the agent queries for past context.
Why try: NetworkChuck connects it via MCP in DW1O6ZBEfDa. Repo moved milla-jovovich/mempalace to MemPalace/mempalace; active. Stores your conversation data locally.
Lets the agent reverse engineer apps and stripped binaries with Ghidra as backend, citing evidence per call.
Why try: Repo description matches the entry's caption word for word. Very active. Dual-use: only analyze software you are authorized to; runs local tooling on untrusted binaries.
Queries app subscription revenue, subscribers and entitlements from RevenueCat.
Why try: Used in DZVXehQRe3T's daily brief. Official RevenueCat hosted server (mcp.revenuecat.ai). Scope the API key read-only unless you want the agent editing products.
Lets the agent read Slack threads (e.g. bug reports) and act on them.
Why try: Claude Code team uses it to fix bugs from Slack (DVQ5o6YDfsp summary). Old reference server is in archived modelcontextprotocol/servers-archived; use Slack's official hosted server. Workspace read access.
claude mcp add taskmaster-ai -- npx -y task-master-ai
Tavily MCP
TryMaintainedofficial
Web search, extract, map and crawl returning clean LLM-ready text instead of blue links.
Why try: Listed in DWncXNCD07J's five-server shortlist. Official Tavily server, active. Paid API key after free tier; fetched pages are untrusted input (prompt-injection surface).
claude mcp add --transport http tavily https://mcp.tavily.com/mcp
Facebook Marketplace MCP
SkipSlow updatescommunity
Searches Marketplace listings on a schedule so the agent can flag below-market deals.
Why skip: Entry says 'one of these' tools, no repo; top match jdcodes1/facebook-marketplace-mcp uses your Chrome session cookies, unofficial scraping (account-ban risk), last commit April 2026, no license.
Searches job listings across Indeed, LinkedIn and Google Jobs for the agent to score and draft proposals.
Why skip: Entry names no repo; several community wrappers exist, borgius/jobspy-mcp-server is most-starred (120), last commit March 2026, no license. Unofficial scraping of job boards.
In May 2025 the Model Context Protocol project moved most of its original reference servers into an archived repository, and they get no further fixes [3]. That set includes the Slack, GitHub, Puppeteer, Brave Search, Google Drive and Postgres servers that many older lists still recommend. The live reference repo now keeps a small set of maintained examples [4]. Prefer each vendor's own server where one exists.
Where to start
For coding work, add Playwright and Context7 first: one lets the agent see and test what it builds, and the other stops it from guessing at library APIs. Add n8n-MCP if you build n8n workflows, and Figma's server if you turn designs into code. Add account-level servers such as Meta Ads, Gmail or Buffer only for a task that needs them, with the narrowest permissions the vendor offers.
Each server you connect adds its tool descriptions to the agent's context on every turn. Remove the ones you do not use. One S-tier post names idle MCP servers, old skills and a bloated CLAUDE.md as the usual places tokens leak (Claude — Burning tokens for no reason is a skill issue.).
What we could not verify
Meta Ads: Meta announced an official MCP server in July 2026, but the entry that covers it does not show which server the creator connected, so it may be a third-party one.
JobSpy and Facebook Marketplace: no entry names a repo, so we list the most-starred matching server and say so in its reason.
TradingView: several unofficial servers compete and none is named in the entry.
REA's GitHub star count looks wrong, so we leave it out.