Best MCP servers in November 2026: 12 official servers scored, and which assistants can actually connect them

Disclosure: NOXI publishes this page and is one of the servers scored. That's why the rules, weights and the source of every cell are printed, along with the servers that finish above NOXI and the one rule where NOXI scores 8 of 15.

Vercel is first with 95. Canva and NOXI both score 93, and the printed tie break places Canva second. Seven servers then finish level on 88: Atlassian Rovo, Cloudflare, GitHub, Linear, Notion, Stripe and Supabase. Figma takes 86 and Sentry 81. NOXI is third of twelve, an ordinary place in a band this tight. GitHub, Atlassian and Notion expose far more tools from a single endpoint than NOXI does, and these rules do not measure breadth.

Most shortlists for this search are lists of repositories with no plan gates in them. This one prints every address, the plan each vendor states, how it authenticates, and whether the official registry holds a record. The surprise is not in the servers: a free Claude account may hold exactly one custom connector, and full MCP support in ChatGPT starts at Business.

Best MCP servers: the 12 at a glance

On seven documented rules checked on 1 November 2026, Vercel scores 95, Canva 93, NOXI 93 and seven more servers 88, so the whole scored field sits inside 14 points.

Where NOXI sits. NOXI is third of twelve, two points behind Vercel and level on points with Canva, which the tie break places above it. NOXI takes full marks on five rules: the published remote address, write actions, a documented OAuth flow, the assistants it names and the registry record. The one rule where it scores 8 rather than 15 is the free-account rule, because the page that publishes the address says nothing at all about plans, and the same silence costs seven other servers the same seven points.

Who should pick what

Five common jobs point to five different servers, each for one documented reason.

• If you want the server with the clearest plan statement → Vercel, because its docs say "Vercel MCP is available on all plans" and describe it as "a remote MCP with OAuth" [5].
• If the recurring chore is code review, issues and pull requests → GitHub, because its install pages name Claude Code, Claude Desktop, Codex, Cursor, VS Code, Visual Studio, Windsurf, Zed and JetBrains [15].
• If the chore is notes, docs and a team wiki → Notion, because its own page calls it "a remote MCP server hosted by Notion" and documents "Create pages and databases." [7]
• If the chore is tickets across Jira and Confluence → Atlassian Rovo, because one endpoint can "Create and update work items, pages, and other content using natural language commands." [9]
• If the chore is turning a written summary into a deck you can present → NOXI, because one address serves Claude, ChatGPT, Cursor, Claude Code and Codex, and the consent screen lists creating, editing and exporting presentations [16].

Key numbers

What an MCP server is, and what changed in the spec

The current specification revision is 2026-07-28, and it keeps exactly two standard transports, so a server you can add to a hosted assistant is always an HTTPS address.

An MCP server is a small service that exposes a product's actions as tools an assistant can call: read this page, open that pull request, create this deck. The Model Context Protocol is the shared shape of those calls, so one server works in every client that speaks it. For the step-by-step of adding one, our guide to Claude connectors, ChatGPT plugins and MCP walks through the settings screens.

Two of those lines matter when you read someone else's shortlist. A tutorial pointing your client at an SSE endpoint is describing a transport the specification has deprecated. And a server with no authorization is not broken: the specification allows it, which is why our auth rule rewards vendors who document a flow.

How we scored: seven rules for an MCP server

Each server is scored on seven plain-point rules totalling 100, applied identically to all twelve, with every cell taken from the vendor's own page opened on 1 November 2026.

Changed recently (vendor pages, checked 1 November 2026). Anthropic's remote MCP help page was updated on 31 October and still prints the one-connector limit for free accounts; ChatGPT's custom MCP page was updated about 27 October, dropping the Developer Mode step and adding a Secure MCP Tunnel route for private servers; the official registry shipped v1.8.7 on 30 October; Cloudflare's server list was updated on 27 October; NOXI's registry record was published on 22 October, the same day Linear's record was updated.

Four readings, printed so the table can be checked

• M3 15 needs the vendor's own page to name the authorization method (OAuth, OAuth 2.0 or OAuth 2.1) or to document an authorize and consent step by that name. M3 8 is a bearer token or API key only, or a browser sign-in documented without the method being named. M3 0 is silence.
• M4 15 needs the vendor to state that no paid plan is required and to gate no documented tool behind a plan or seat. Where the server is free but named tools need a paid plan, licence or seat, M4 is 8. Where the vendor says nothing about plans, M4 is 8.
• M6 is decided by what the official registry returns today, not by a vendor's claim. A record carries every version its vendor has published, so the current one is the one that counts, and a search that does not ask for it returns the oldest first. GitHub's separate "GitHub MCP Registry" is another index and does not count.
• M7 10 needs the page to show one complete task: a worked prompt and its result, or a named round trip. A list of tools, or "access to our product", is 5. This rewards documentation, not capability, and the section below says so.

Which servers are in, and which are only mentioned

Twelve servers are scored: a vendor's own server for a product people use at work, reachable over the public internet. Local-only servers are described but not ranked, because M1 would cap them at 10 by construction. Playwright runs through npx @playwright/mcp@latest, and its page warns that "Playwright MCP is not a security boundary." Chrome DevTools runs the same way and does hold a registry record. Hugging Face, Asana, PayPal and Zapier are remote and real but sit outside the twelve. Slack publishes an official server at mcp.slack.com/mcp with OAuth and writes, and is named here only.

Every cell, every address, and the arithmetic

Vercel's 95 is 20 plus 15 plus 15 plus 15 plus 15 plus 10 plus 5, and every other total is built the same way.

The two ties, decided by the printed rule

• Canva 93 and NOXI 93. Level on M1, M2 and M5, and level on every other cell too, so the table's printed order decides. Canva comes before NOXI, so Canva is second and NOXI third.
• Seven rows tie at 88, level on M1, M2 and M5 as well, so they are listed alphabetically, fourth to tenth: Atlassian Rovo, Cloudflare, GitHub, Linear, Notion, Stripe and Supabase. The tie is the finding, not a flaw. On documented capability these seven are identical cell by cell, which is why the useful question is about your assistant rather than the servers.

Where the gaps come from

• Vercel minus NOXI = 2. Vercel is 7 ahead on the free-account rule and 5 behind on the finished-job rule. Check: 7 minus 5 = 2.
• Canva minus NOXI = 0. The rows are identical cell by cell, which is why the tie break had to decide.
• NOXI minus any of the seven at 88 = 5, all on the finished-job rule: those seven list their tools, NOXI prints four worked prompts.
• The seven at 88 minus Sentry = 7, all on the write rule. Sentry is last because it is the only row whose page names no create or update action.

How sensitive the order is

• NOXI 100 and first if the product's own pricing counted as the connector's plan statement and M4 became 15. The table scores 8 because the page publishing the address says nothing about plans, and that silence costs Notion, Linear, Atlassian Rovo, Supabase, Stripe, Cloudflare and Sentry the same seven points.
• Figma 93 if a browser sign-in and allow-access flow counted as an OAuth flow without the word OAuth (M3 15). Figma's installation page gives numbered steps but never names the method.
• Sentry 88 if "managing projects" counted as a documented create or update action (M2 15).
• Cloudflare 81 if "access to the entire Cloudflare API" is not read as a documented write action (M2 8).
• Vercel 88 if "Vercel MCP is available on all plans" is read as covering the server but not every tool (M4 8), which would put Canva first on 93.
• Every scored row at M7 10, if the finished-job rule were read as capability rather than documentation: Vercel 100, then nine servers on 93 including NOXI, with Figma and Sentry on 86. NOXI's place does not improve under that reading, which is worth saying out loud.

Only two readings move the top of the table: counting a product's general pricing as the server's plan statement puts NOXI first at 100, and reading Vercel's plan line narrowly puts Canva first at 93.

Which assistants can connect a remote server, and on which plan

Claude allows custom connectors on every plan and limits a free account to one, while ChatGPT reserves full MCP support for Business and Enterprise or Edu customers on the web, so the same twelve servers behave very differently depending on where you type.

Two things get mixed up here, so keep them apart: adding your own server, and using a connector a vendor published for a surface. Canva's help centre names 21 assistants and tools including Copilot and Gemini, and Adobe's Express pricing page says its product is "Integrated into Claude, ChatGPT, Microsoft Copilot, and Gemini via add-on connectors". Neither is the same as pasting your own address into a consumer assistant, and that route does not exist there.

One more detail: OpenAI's page says "ChatGPT uses a \"frozen\" snapshot of its available tools and inputs", so a tool a vendor shipped last week may not appear until that snapshot refreshes. If you work in a chat window rather than an editor, our ranking of the best AI tools for work scores the surfaces themselves.

The servers that can write, not just read

Eleven of the twelve scored servers document create or update actions, which is the biggest difference between a server that saves time and one that only answers questions.

A write action is also the one to think hardest about. Stripe touches money objects, Supabase runs migrations, Cloudflare sits at the network edge, and none of these rules measures how safe a write is in production. That review belongs to your own team.

Auth: what OAuth 2.1 means for you here

Eleven of the twelve scored servers document an OAuth flow on their own page, so this rule separates exactly one row: Figma's 8 is the difference between its 86 and the 93 it would otherwise score.

The flow is the same everywhere. You paste the address, a browser window opens on the vendor's domain, you sign in there, and a consent screen lists what the assistant may do. The token that comes back is bound to that one server, which is what RFC 8707 resource indicators are for, and you revoke it where you created it.

Figma shows what this rule actually measures. The flow is documented step by step, with a browser window and an allow-access screen, and anyone following it will connect. What is missing is the word itself. Read it the other way and Figma scores 93, which the sensitivity list prints.

The official registry, and why "in preview" matters

All twelve scored servers hold a record in the official registry at registry.modelcontextprotocol.io, which says on its own page that it "is currently in preview. Breaking changes or data resets may occur".

The registry is the closest thing MCP has to an index: one place where a vendor publishes a name, a version, a website and the remote address, under "Model Context Protocol a Series of LF Projects, LLC". It is not an app store. No review, no badge, no ranking, so a record means a vendor published one and nothing more. Hence 10 points, not 20.

Ask the registry for the current version, or it will mislead you. A record keeps every version its vendor has ever published, and a plain keyword search hands all of them back oldest first. GitHub's record is the clearest example: the 2025 versions carry only a container image for a local install over stdio, while the current one, 2.0.3, published 21 October 2026 and marked active, carries the remote address scored here, api.githubcopilot.com/mcp/, over Streamable HTTP. Add version=latest to the query and you get the second answer. We read every row in this table that way on 1 November 2026.

Two more things worth knowing. The registry's about page names GitHub among "major trusted contributors to the MCP ecosystem such as Anthropic, GitHub, PulseMCP, and Microsoft", and GitHub also runs a "GitHub MCP Registry" of its own, which is a separate index and earns nothing under this rule. Asana and Slack are the two servers here with no record at all, although both publish an official remote server.

What NOXI's record says, quoted from the registry API rather than from any claim: name ai.noxi/noxi, title "NOXI", version 0.1.0, description "Free AI presentation maker: create and edit decks, export PowerPoint and PDF, Arabic right-to-left", website noxi.ai/connect, one remote of type streamable-http at mcp.noxi.ai/mcp, status active, isLatest true, published 2026-10-22T14:25:06Z. Listed, in a registry still in preview. Not featured, not approved and not verified, because the registry does none of those things.

NOXI: third of twelve, and what its cells say

NOXI scores 93 and finishes third, two points behind Vercel and level with Canva, with full marks on five of the seven rules.

The honest comparison is about breadth. GitHub, Atlassian and Notion expose far more tools from one endpoint, because they sit in front of whole platforms, while this server covers one product and one job. These rules do not measure that, which is the first line of the section below.

Three more lines from the same page: "Every assistant connects to the same address. You sign in to NOXI in your browser, so there is no API key to copy."; "A connection stays active for up to a year while you use it at least once a month."; and "NOXI only works with presentations in the account you connected." The Claude path is the one Anthropic documents: "In Claude, open Customize > Connectors, select + Add, then Add custom connector." If the job is slides rather than servers in general, our scored page on MCP servers for PowerPoint is the specialist version of this one, and the best AI for PPT ranks the deck makers.

What changed this week

Five dated changes landed on official pages in the seven days to 1 November 2026, and not one of them was a new specification revision.

• 31 October 2026. Anthropic's remote MCP help page carries "Updated yesterday" and still prints "Free users are limited to one custom connector.", as does the developer documentation.
• About 27 October 2026. ChatGPT's custom MCP help page carries "Updated: 5 days ago". Adding a server no longer needs Developer Mode first, and a Secure MCP Tunnel route now appears for servers you keep off the public internet.
• 30 October 2026. The official registry shipped v1.8.7, after v1.8.6 on 27 October made search terms need three or more characters and v1.8.5 on 26 October added performance diagnostics. It is still in preview.
• 28 October 2026. The VS Code MCP servers documentation and the Microsoft 365 Copilot plugin MCP page both carry that date.
• 27 October 2026. Cloudflare's list of its own MCP servers was updated. On 22 October, NOXI's registry record was published and Linear's record was updated.

The specification itself is unchanged: 2026-07-28 is still the current revision.

What these rules don't measure

Seven rules cannot describe a server you will live with every day, and the gaps are large enough that a reader should know them before copying the order above.

• How many tools each server exposes. GitHub and Atlassian lead by a wide margin. Atlassian's single endpoint reaches Jira, Confluence, Jira Service Management, Bitbucket Cloud, Loom, Goals, Teams, Teamwork Graph and code search, and Cloudflare's covers "over 2,500 endpoints". Breadth earns nothing here.
• Latency, rate limits and uptime history. Only Figma publishes a per-plan rate table, so there is nothing comparable for the rest.
• How safe a write action is in production. This matters most for Stripe, Supabase and Cloudflare, and it is your team's judgement about your own data.
• Self-hosted and local servers. That is the whole of Playwright and Chrome DevTools, and much of the community ecosystem.
• Community servers. Thousands exist, many are excellent, and none is scored here, because the rules ask what a vendor documents about its own product.
• Maturity. A record from September 2025 and one from last week both score 10, although one has had a year of fixes.
• The review burden on your team, and whether your employer allows third-party connectors at all. In a managed workspace that decision is not yours.
• How well a vendor documents a worked example, which is what the finished-job rule actually rewards. GitHub finishes far more jobs than its 5 suggests.

What they are good for is narrower and still useful: whether a server exists at a published address, whether it can change things rather than only read them, how it authenticates, what the vendor says about plans, which clients it names, and whether a record exists for it. That is the part most lists leave out.

Worked example: one connector, chosen properly

Example Daniel Okonkwo is a fictional solo developer in Lagos with three retained clients. He uses Claude on the free plan, so he may add exactly one custom connector. Everything below follows from that one number.

His arithmetic is simple. Three clients, one sprint review each, three decks a month. One deck by hand takes him about 80 minutes, so 3 times 80 = 240 minutes, which is 4 hours. Check: 240 divided by 60 = 4. Code review is already fast, because he lives in the editor, where a local server costs no connector slot.

His decision. The deck chore gets the slot, because it is the one he does badly and resents. GitHub moves into his editor, where Claude Code and Cursor reach it without touching the Claude connector limit. Sentry stays a browser tab. He sets a review date of 1 February 2027.

The lesson is not which server he picked. It is that the free limit, not the quality of any server, decided. On a paid plan the problem is the opposite: too many good options, no slot to ration.

Build it: make a deck inside ChatGPT with the NOXI plugin

If the chore in your week is the deck, the easiest route is one you never configure: NOXI is listed in ChatGPT's plugin directory, so you install it from the sidebar instead of pasting an address.

Installing or using a plugin "depends on your plan, workspace, role, region", in OpenAI's own words, so check what your account allows first. To add the server by hand instead, the same address, mcp.noxi.ai/mcp, is the one you paste into Claude, Cursor, Claude Code or Codex from noxi.ai/connect. The launch write-up for the NOXI plugin in ChatGPT has the rest.

Decks started from a brief are researched on the web first, and slides with figures carry a source line. Attach a file instead and research is off by default, so put the numbers you need in the brief. Our comparison of DeepSeek and ChatGPT covers the chat side, and our ranking of presentation software the editors.

Frequently asked questions

What are the best MCP servers right now?

On seven printed rules checked on 1 November 2026, Vercel scores 95 of 100, and Canva and NOXI take 93 each. Seven servers then tie on 88: Atlassian Rovo, Cloudflare, GitHub, Linear, Notion, Stripe and Supabase. Figma takes 86 and Sentry 81. All twelve are official vendor servers with a published address.

What is an MCP server, in one paragraph?

An MCP server exposes one product's actions as tools an assistant can call, using the Model Context Protocol so the same server works in every client that speaks it. Remote servers are an HTTPS address you paste in; local servers run on your own machine over stdio. The current specification revision is 2026-07-28.

Which MCP servers work with Claude?

All twelve scored here. Anthropic's help page, updated 31 October 2026, says custom connectors using remote MCP work on Free, Pro, Max, Ultra, Team and Enterprise plans, and that free accounts may add one. Claude calls your server from Anthropic's cloud, so it must be reachable over the public internet.

Can I use MCP servers in ChatGPT?

Partly. OpenAI's help page, updated about 27 October 2026, says custom MCP server connections and full MCP support are for Business and Enterprise or Edu customers on ChatGPT web. Pro accounts get read and fetch permissions only, and the answer on mobile is "No - web only". Free and Plus are not addressed.

Does Microsoft Copilot support MCP?

Not as a route for a server you host yourself. No Microsoft or Google first-party page lets an individual add a custom remote server to consumer Copilot or the Gemini app. The documented routes are Microsoft 365 Copilot plugins, Copilot Studio and Gemini Enterprise, which sits under pre-GA terms. Checked 1 November 2026.

How many connectors can a free Claude account have?

One. Anthropic's remote MCP page, updated 31 October 2026, states that free users are limited to one custom connector, and the developer documentation repeats it. Paid plans from Pro upwards are not capped at one, so the limit shapes the choice only on the free tier. Local servers in an editor are unaffected.

Are official MCP servers free?

Usually the server is, and some tools are not. Only Vercel states plainly that its server needs no paid plan. Canva's is available to all accounts but gates brand template actions behind a paid plan, Supabase charges for branching tools, Atlassian's consume Rovo credits, and GitHub mirrors each feature's own licence. Checked 1 November 2026.

Is SSE still supported?

As a streaming format, yes; as a transport, no. The specification says the server returns either application/json or text/event-stream inside Streamable HTTP. The old two-endpoint HTTP plus SSE transport is deprecated: new implementations should not adopt it and existing ones should migrate. Revision 2026-07-28 keeps exactly two standard transports.

What is the official MCP registry?

It is the index at registry.modelcontextprotocol.io where vendors publish a server's name, version, website and remote address, under Linux Foundation governance. Its own page says it is currently in preview and that breaking changes or data resets may occur. All twelve of the servers scored here hold a record there on 1 November 2026.

Can an MCP server make a PowerPoint?

Yes. NOXI's server, at mcp.noxi.ai/mcp, documents creating and editing presentations and exporting them as PDF or PowerPoint, and it scores 93 of 100 on our seven rules. Canva's server documents a round trip ending in an export too. Both were checked on 1 November 2026.

Sources & methodology

Every page below was opened on 1 November 2026, signed out unless the page required otherwise, with no forms submitted and nothing installed. The registry was read through its public API. Scores are copied from our shared November 2026 score file; quotes are 15 words or fewer. Daniel Okonkwo is fictional, and the hours in his table are his own estimate, not a measurement.

• [1] Model Context Protocol specification and project pages, checked 1 November 2026 · Specification, latest revision; Transports; Streamable HTTP; Authorization; About the registry; Governance.
• [2] Anthropic, custom connectors using remote MCP, page updated 31 October 2026 · Get started with custom connectors using remote MCP; Remote MCP, developer documentation.
• [3] OpenAI, connecting custom MCP servers in ChatGPT, page updated about 27 October 2026 · Connecting custom MCP servers in ChatGPT; Plugins in ChatGPT.
• [4] The official MCP registry, read through its public API on 1 November 2026 · Health endpoint; Servers endpoint.
• [5] Vercel · Vercel MCP, last updated 15 September 2026.
• [6] Canva · Verify your MCP integration; Canva AI Connector setup; AI connectors for Canva.
• [7] Notion · Notion MCP; Get started with MCP.
• [8] Linear · Linear MCP server.
• [9] Atlassian · Rovo MCP server, last updated 2 September 2026.
• [10] Supabase · Model context protocol.
• [11] Stripe · Stripe MCP server.
• [12] Cloudflare · MCP servers for Cloudflare, last updated 27 October 2026.
• [13] Figma · Figma MCP server; Remote server installation; Rate limits and access.
• [14] Sentry · Sentry MCP server.
• [15] GitHub · Set up the GitHub MCP server; Use the GitHub MCP server; Extend Copilot Chat with MCP; github/github-mcp-server.
• [16] NOXI connector page and product documentation · noxi.ai/connect, read 1 November 2026.

Pick the server that removes the chore you resent. If that chore is the deck, connect NOXI to the assistant you already use, or read our guide to building slides with Claude.