Which AI Tools Support MCP in 2026? Full List

The Model Context Protocol stopped being a Claude-only thing in 2025. Every serious AI client now supports it in some form. The catch is that “support” means different things in different places. Claude Desktop runs MCP servers like it invented them, which it did. ChatGPT runs them too, behind a setting most users never find, on plans most users do not have.

The question “which AI tools support MCP” looks simple from outside the question and gets more complicated the closer you look. Free or paid. Read tools or write tools. Local servers or hosted ones. Stable or beta. The honest answer is a tier list, not a yes or no.

This post is that tier list. Every major AI client that supports MCP in May 2026, what tier of support it has, what it costs to use, and the small print that the marketing pages do not put on the front page.

What “MCP support” actually means

The Model Context Protocol is an open standard released by Anthropic in November 2024. It defines a clean separation between two roles. The MCP server is the data source: a process or remote URL that exposes a list of tools the AI can call. The MCP client, also called the host, is the AI tool itself: Claude Desktop, Cursor, ChatGPT, whatever you happen to be typing into.

“Does this tool support MCP” really means “can this tool act as an MCP client.” When you read that Cursor “supports MCP,” it means Cursor knows how to read tools from any MCP server, call them mid-chat, and feed the results back into its reasoning. The server can be anything. A GitHub MCP server. A Postgres MCP server. A bookmark search server like ContextBolt. The protocol does not care.

The reason this matters is portability. The same server works in every supported client without changes. You wire ContextBolt once, then connect Claude Desktop, Cursor, and ChatGPT to it. Three different tools, one source of truth for your bookmarks. For the broader explainer of how MCP works, What is MCP? is the companion read.

Different clients support different slices of the protocol. Some support remote HTTP servers but not local stdio servers. Some support read tools but not write tools. Some support OAuth. Some only support API key auth. The tier list below is sorted by depth of that support, not just by whether the marketing page says the word “MCP” somewhere.

Tier 1: First-class MCP support

These are the clients where MCP is a primary design feature. Anthropic shipped the protocol and these two products at roughly the same time. The integration is the deepest, the docs are the cleanest, and the bug rate is the lowest.

Claude Desktop

The original MCP client. Claude Desktop’s three install paths cover every level of user: Custom Connectors in the UI for remote servers, .mcpb Desktop Extensions for one-click local installs, and the raw claude_desktop_config.json file for full control. Supports local stdio servers, remote HTTP servers, and OAuth-authenticated cloud servers. Available on macOS and Windows. Custom Connectors and full MCP work on all paid plans (Pro, Max, Team, Enterprise). Free users can use Anthropic-built connectors but not arbitrary servers.

This is the consumer-facing default. If your AI work is conversational and you do not live in an IDE, Claude Desktop is the right home. The setup guide lives at Claude Desktop MCP: Setup, Servers, and the 2026 Stack.

Claude Code

The CLI cousin of Claude Desktop. Claude Code runs in the terminal, reads .mcp.json from the project root, and treats every server in there as a tool the agent can call. Three scopes: project-level (.mcp.json, commit to git), user-level (~/.claude.json), and remote (HTTP servers added with claude mcp add).

Claude Code’s MCP support is identical in shape to Claude Desktop but tuned for coding work. Same servers, different config file. Available on all paid Anthropic plans and free on the Pro plan with usage limits.

Tier 2: Strong native support

These clients did not invent MCP, but they shipped support early, kept up with the spec, and treat it as a feature you should know about, not a hidden setting. Day-to-day, the experience is indistinguishable from Tier 1.

Cursor

Cursor added MCP in late 2025 and the support is now a first-class part of the editor. Config lives in ~/.cursor/mcp.json for user-level servers and .cursor/mcp.json for project-level. Supports stdio, SSE, and HTTP servers. The Cursor agent reads available tools at the start of every Composer session and picks which to call as it works.

Cursor’s MCP support is free on the Hobby plan. Full setup guide at Cursor MCP: Setup, Servers, and the 2026 Stack.

Windsurf

Codeium’s Windsurf editor shipped MCP support for Cascade (the agent) in early 2025. Config lives at ~/.codeium/windsurf/mcp_config.json. Supports the standard stdio and HTTP transports. The 100-tool ceiling is the main quirk to watch: too many servers and the agent starts thrashing. Setup details at Windsurf MCP: Setup, Servers, and the 2026 Stack.

VS Code with GitHub Copilot

VS Code’s MCP support shipped to all users in v1.102 in July 2025 alongside Agent mode going generally available, per the Visual Studio team’s announcement. Two config locations: workspace at .vscode/mcp.json and a user profile that syncs via Settings Sync.

The biggest VS Code-specific gotcha: the root key is servers, not mcpServers, and tools only run when Copilot Chat is switched to Agent mode. Otherwise nothing fires. Also features the largest curated MCP gallery, accessible from the Extensions panel with the @mcp filter. Detail at VS Code MCP: Setup, Servers, and the 2026 Stack.

Tier 3: Behind a toggle, a paid plan, or a beta

These clients support MCP, but the support is gated. Sometimes by a hidden setting. Sometimes by a paid plan. Sometimes by both. The protocol works. The path to the toggle is the friction.

ChatGPT (Developer Mode)

OpenAI added full MCP support to ChatGPT in late 2025 under Developer Mode. Toggle it on at Settings, Connectors, Advanced. Once on, you can add any compatible MCP server, not just the ones in OpenAI’s curated Apps panel.

The catch list is unusually long. Developer Mode is paid plans only (Plus, Pro, Team, Business, Enterprise). Write tools are gated to Business and Enterprise. Read tools work on Plus and up. The Apps panel itself, which is a different feature, is unavailable in the EEA and UK as of May 2026. Pop your bookmarks server in via Developer Mode and ChatGPT can search your saved tweets mid-conversation. Walkthrough at Make ChatGPT Remember Your Notes: 5 Methods (2026).

Zed

Zed added MCP support in 2025 via its Assistant panel. Config lives in ~/.config/zed/settings.json under the context_servers key, which is Zed’s name for MCP servers. Supports the standard transports. Lighter than Cursor or VS Code but fast, and the Vim mode is the best in any AI-first editor.

Cline (and the forks)

Cline is the open-source coding agent that runs inside VS Code as an extension. Full MCP support, configurable via the Cline UI or a settings file. The forks (Roo Code, Kilo Code, others) inherit MCP support from the parent codebase. Free, bring your own API key.

Continue

Continue is the other major open-source coding agent extension for VS Code and JetBrains. MCP support landed in 2025 and the configuration lives in the ~/.continue/config.json file. Same shape as the other JSON-driven clients.

LibreChat

LibreChat is the most popular open-source ChatGPT-style web UI. It supports running multiple models and added MCP server support in 2025. The configuration is YAML-based and lives in the LibreChat config file. Useful for teams that want a self-hosted multi-model chat with MCP, without paying ChatGPT Team prices.

Cherry Studio

Cherry Studio is a desktop chat client that supports multiple model providers and ships with MCP support. Configuration is UI-driven, which makes it the closest cousin to Claude Desktop for users who want to chat with OpenAI, Anthropic, and local models from one app.

Tier 4: Not supported yet, or barely

The list of holdouts shrinks every quarter. As of May 2026, the obvious gaps are:

• Google Gemini. Public Gemini products do not expose user-configurable MCP yet. Vertex AI on Google Cloud supports the protocol for developer use, but the consumer Gemini app and Gemini for Workspace do not.
• Microsoft Copilot for consumers. GitHub Copilot is fine (covered above). Microsoft 365 Copilot has its own plugin system (Copilot Studio, agents) but does not act as a general MCP client. Microsoft has signaled support is coming, with no shipped feature yet.
• Perplexity. Browsing AI with sources, no general MCP client. Spaces exist for fixed reference material, in the same shape as Custom GPTs and Projects.

The pattern is: tools built by Anthropic and tools built by the editor crowd shipped MCP early. Tools built by general consumer AI companies have been slower, and most are still gated by paid plans where they exist at all.

How MCP support varies across clients

The “supports MCP” line on every tool’s marketing page hides a lot of variation. The table below covers the things that actually break setups.

Three patterns fall out of the table. The Anthropic-built clients are the easiest to live with. ChatGPT is the only major client that does not support local stdio servers, which makes sense (you cannot run a process on OpenAI’s servers). The free tier story matters more than people realize: if you are not yet on a paid plan, the list of clients you can actually use with MCP shortens to about five.

How to pick the right MCP host

The honest answer: pick the client you would have used anyway, and check this list once to confirm it supports MCP. Forcing a workflow shift to chase deeper MCP support is rarely worth it.

A few specific recommendations.

If you live in a terminal, use Claude Code. Best CLI agent on the market and the MCP support is exactly what you’d expect from the company that wrote the spec.

If you do conversational work outside a coding context, use Claude Desktop or ChatGPT. Claude Desktop has cleaner MCP wiring. ChatGPT has the larger Apps panel and the wider model spread. If you are paying for one of them already, that is the right one to start with.

If you live in an editor, your existing editor is probably fine. Cursor, Windsurf, and VS Code all have strong MCP support. The differentiator at this point is the editor itself, not the protocol. Cursor and Windsurf for AI-first workflows. VS Code with Copilot for everything else.

If you want a self-hosted multi-model setup, use LibreChat or Cherry Studio. Both keep your data local and both speak MCP. Trade-off: more setup, no managed updates.

The one piece of advice everyone underrates: pick a single MCP server stack first, then pick clients around it. The protocol’s whole point is portability. A bookmark MCP, a GitHub MCP, a Postgres MCP, and a filesystem MCP is a useful four-server starter stack in every client above. Pick the stack, then your editor and your chat app become interchangeable.

Where ContextBolt slots into this picture

A quick honest note since you are reading this on the ContextBolt blog.

ContextBolt is a Chrome extension that captures bookmarks from X, Reddit, and LinkedIn, auto-tags them by topic, and runs semantic search over them. The free tier ($0, 150 bookmarks) gives you AI tagging, topic clustering, and search inside the extension. Pro ($6 a month) adds unlimited bookmarks, encrypted cloud sync, and a personal MCP endpoint.

The MCP endpoint is the part that matters for this article. Wire it into any Tier 1, 2, or 3 client above and your saved content becomes a tool that AI can call. Ask Claude what you saved about LLM evals last month and it calls search_bookmarks against your collection. Same endpoint, every client. Start in Claude Code, switch to ChatGPT next year, the bookmark layer follows you. The end-to-end setup for Claude Code is at Add Your Bookmarks to Claude Code via MCP.

The honest take

The protocol won.

In December 2024, MCP was an Anthropic side project that the rest of the industry could have ignored. By mid-2025, every coding editor of consequence had shipped support. By late 2025, OpenAI shipped it too. By mid-2026, asking whether your AI tool supports MCP is the same question as asking whether your laptop has Wi-Fi. The yes is implied. The interesting question is the small print.

The small print is the part most lists skip. Free plans are still locked out of the deepest features. ChatGPT gates write tools behind Business. Anthropic gates user-configurable connectors to paid plans. The protocol is universal, the access model is not.

The other part the marketing pages do not say out loud: the hard part of an MCP stack is not adding clients, it is picking the right servers. Most users wire two or three, hit “tool overload” when the agent gets confused, and quietly turn the rest off. The right move is the opposite: pick a small, deliberate server stack that covers the data you actually use, then add clients in front of it as your workflow changes.

The clients are a commodity. Pick the servers carefully and the clients become interchangeable.