Best Remote MCP Servers in 2026: The Working List

TL;DR - Remote MCP servers connect to Claude, ChatGPT, and Cursor with just a URL — no install, no config file, no Node runtime. - The ecosystem has consolidated around official first-party servers: GitHub, Notion, Linear, Sentry, Stripe, Atlassian, Asana, Cloudflare and others all run their own hosted endpoints now. - Most remote servers are API wrappers: they expose one product's data to your model. A newer class — including OpenHelm's five servers — executes entire tasks in a sandbox and returns finished work. - OAuth is now the default auth pattern; if a server still asks you to paste a long-lived API key into a config file, treat that as a smell. - This list only includes servers we could verify are live in July 2026.

---

What Counts as a "Remote" MCP Server

The Model Context Protocol supports two fundamentally different deployment shapes. A local server is a process on your machine — usually spawned over stdio by Claude Desktop or Cursor, configured in a JSON file, dependent on whatever runtime it needs. A remote server is a hosted HTTPS endpoint. You paste a URL into your client, complete an OAuth flow, and you're connected.

Through 2025 the ecosystem was dominated by local servers, because that's what the first wave of tooling supported. That flipped. ChatGPT only supports remote servers. Claude's connectors directory is built around one-click remote connections. Most major SaaS vendors now run an official hosted endpoint rather than pointing users at a GitHub repo. If you're choosing servers to rely on in 2026, remote-first is the sensible default — we cover the trade-offs in detail in remote vs local MCP servers.

A note on method: MCP directories list tens of thousands of servers, most of which are abandoned weekend projects. Everything below is an official first-party server or a maintained hosted product, verified live as of July 2026. Where we state an endpoint URL, it's the documented one; where a vendor doesn't publish a stable URL, we say so.

---

The Two Kinds of Remote Server (Read This First)

Before the list, one distinction that matters more than any individual entry.

API-wrapper servers expose a single product's API as tools. The GitHub server gives your model list_pull_requests; the Linear server gives it create_issue. The model does the thinking; the server does one call at a time. These are excellent for giving an assistant hands inside a product you already use.

Task-execution servers accept a goal and do the work elsewhere. The tool call returns a run handle; an agent executes in isolated infrastructure — browsing, reading sources, composing output — and you poll for the finished result. The model that called the tool doesn't need to orchestrate 40 sub-steps or hold every intermediate page in context.

Most of this list is the first kind, because that's most of the ecosystem. The last section covers the second kind, which is where OpenHelm's servers sit. Neither is "better" — they solve different problems, and a good setup usually connects both.

---

Best Remote MCP Servers for Developers

GitHub (official)

• Endpoint: https://api.githubcopilot.com/mcp/
• Auth: OAuth (one-click in VS Code 1.101+) or a personal access token

The reference example of a first-party remote server. Issues, pull requests, code search, Actions runs, security alerts — all exposed as tools, hosted by GitHub, updated without you touching anything. If your workflow lives on GitHub, connect this first. The main caveat: it's a big tool surface, and models occasionally pick the wrong tool on ambiguous prompts, so scope your requests ("in repo X…").

Sentry (official)

• Auth: OAuth against your Sentry org, inheriting its existing permission model

Query issues, stack traces, and release health from inside your editor or chat. Particularly good paired with a coding agent: "look up the top unresolved issue in project X and propose a fix" becomes one conversation instead of four tabs.

Cloudflare (official)

Cloudflare runs a *suite* of remote servers rather than one — documentation, Workers bindings, and observability each get their own endpoint, documented in the Cloudflare docs. Notable because Cloudflare also published the open-source scaffolding much of the remote-MCP ecosystem builds on (OAuth provider libraries, deploy-to-Workers templates).

DeepWiki (Devin/Cognition)

• Endpoint: https://mcp.deepwiki.com/mcp (Streamable HTTP) or /sse
• Auth: none — free and open

Ask questions about any public GitHub repository and get answers grounded in an AI-generated wiki of that codebase. One of the few genuinely useful no-auth servers, and a good first server for testing that your client's remote MCP support works at all.

---

Best Remote MCP Servers for Project & Knowledge Work

Notion (official)

• Endpoint: https://mcp.notion.com/mcp
• Auth: OAuth

Search, read, and write pages and databases. Notion's hosted server replaced the earlier self-hosted notion-mcp-server as the recommended path, and it's noticeably better at returning results in a token-efficient format than the raw API wrapper was.

Linear (official)

• Endpoint: https://mcp.linear.app/mcp
• Auth: OAuth

Create and update issues, manage projects, query cycles. Linear shipped remote MCP early and it shows — tool descriptions are tight, and agents use it reliably. If your team runs on Linear, this plus the GitHub server covers most of a developer's coordination surface.

Atlassian (official)

• Auth: OAuth

Jira and Confluence access through Atlassian's hosted endpoint. More enterprise-shaped than Linear's (permissions, site selection), but it means agents can finally read the Confluence page where the actual requirements live.

Asana (official)

• Endpoint: https://mcp.asana.com/sse
• Auth: OAuth

Tasks, projects, and portfolios. Note the SSE transport — older-style but still supported by every major client.

---

Best Remote MCP Servers for Business Operations

Stripe (official)

• Endpoint: https://mcp.stripe.com
• Auth: OAuth

Query customers, payments, subscriptions, and invoices, plus search Stripe's documentation. Treat write-capable tools here with respect: an agent with refund powers should be running under human-in-the-loop controls, not full autonomy.

PayPal, Square (Block), Intercom, HubSpot (official)

All four run official OAuth-based remote servers — payments and invoicing for PayPal and Square, customer conversations for Intercom, CRM objects for HubSpot. Endpoints are in each vendor's developer docs. The pattern across all of them: read tools are safe to hand any agent; write tools deserve an approval gate.

---

Task-Execution Servers: When You Want Finished Work, Not API Calls

Everything above makes your model better at *operating software you already use*. The gap they don't cover: work that takes twenty minutes of browsing, reading, and synthesis. A chat model calling search fifteen times in a row is slow, expensive in context, and fragile.

OpenHelm runs five remote MCP servers at mcp.openhelm.ai built on a different contract: a tool call starts an asynchronous agent run in an isolated cloud sandbox with a real browser, and returns a handle you poll with check_result. The MCP layer itself is stateless and stores no secrets.

All five use OAuth 2.1 — you never paste an API key into a client config — and connect to ChatGPT, Claude, and Cursor by URL, or via npx mcp-remote for clients that need a local bridge. Honest positioning: if you just need your assistant to create a Linear issue, use Linear's server — it's simpler and instant. OpenHelm's servers earn their place when the task is a *job* rather than a *call*: research that reads dozens of sources, outreach that runs a whole conversation, a code change that needs tests run before the PR opens. Setup details are in the MCP docs.

---

More Official Servers Worth Knowing

Beyond the entries above, the official-remote roster keeps growing. Servers verified live from their vendors in 2026 include Slack (messages and channels), Figma (design file context for coding agents), Vercel and Neon (deployments and Postgres respectively — both shipped remote endpoints this year), Supabase, and Hugging Face (model and dataset search). We haven't given each a full entry because the evaluation is identical: it's the vendor's own hosted endpoint, OAuth-authenticated, documented in their developer docs. If you use the product, the official server is almost always the right pick over any community alternative — it tracks API changes the day they ship and inherits the vendor's permission model.

Two practical notes from running a stack of these side by side. First, tool-count bloat is real: connect eight servers with thirty tools each and models start making slower, worse tool choices. Most clients let you enable a subset of tools per server — do it, aggressively. Second, watch for overlapping capabilities: if both your GitHub server and a task-execution server can open PRs, tell the model which to use in your prompt or system instructions, or you'll get nondeterministic routing between them.

---

How to Choose From This List

1. Start with the products you already pay for. An official server for a tool your team lives in beats any generic server.
2. Prefer OAuth over pasted keys. Every server above supports it. Long-lived keys in client configs get leaked.
3. Separate read from write. Connect read-heavy servers freely; put write-capable tools (payments, email, code) behind approval or scoped accounts.
4. Add one task-execution server for the long-running work chat loops handle badly.
5. Check the directory before building your own. The MCP registry and directory ecosystem makes it easy to verify whether an official server already exists.

---

Frequently Asked Questions

What is the best MCP server overall?

There's no single answer because servers aren't interchangeable — each exposes different capabilities. For most developers, GitHub's official remote server is the highest-value first connection. For teams, whichever official server matches your project tool (Linear, Notion, Atlassian). For delegated multi-step work, a task-execution server like OpenHelm's general-purpose MCP.

Which MCP servers work with Claude?

All of the servers in this list. Claude (web and desktop) supports remote MCP servers via the connectors directory and custom connector URLs, and Claude Desktop additionally supports local stdio servers. See our Claude MCP setup guide for the walkthrough.

Do remote MCP servers work with ChatGPT?

Yes — ChatGPT supports remote HTTP/SSE MCP servers (it does not support local stdio servers). Every server listed here is remote and therefore ChatGPT-compatible; the connection steps are covered in how to use an MCP server with ChatGPT.

Are remote MCP servers safe to connect?

The auth model is the main risk surface. OAuth-based servers (everything in this list) mean the server holds a scoped, revocable grant rather than you distributing raw API keys. Beyond auth, apply least privilege: connect read-only where possible, and gate write-capable tools behind approvals.

How many MCP servers exist in 2026?

Community directories index tens of thousands (PulseMCP tracks 20,000+; Glama indexes 50,000+ open-source servers), but the number of *maintained, official, remote* servers is far smaller — dozens, not thousands. That's why curated lists like this one exist.