Research reportai-agents
What Is MCP, and Which AI Agents Actually Support It?
MCP is becoming the standard way agents connect to tools and data. We checked our 45-tool index for documented support — near-universal except sales agents.
Every third listing page in this space now name-drops “MCP support” as a feature, with no explanation of what it actually buys a buyer. Here’s the plain definition, why it matters when you’re picking an agent to build on, and — because a definition alone isn’t worth publishing — a real count of which tools in our own 45-listing index document real MCP support today, and which don’t yet.
What MCP actually is
The Model Context Protocol (MCP) is an open standard, introduced by Anthropic on 2024-11-25 (then open-sourced the same day), for connecting AI applications to external data sources and tools through one common interface instead of a custom integration per pair. Anthropic’s own framing is the clearest: before MCP, “every new data source requires its own custom implementation, making truly connected systems difficult to scale” — the N-times-M problem, where N different AI applications each needed a bespoke connector to M different tools and data sources. MCP replaces that grid with one protocol both sides implement once.
The protocol’s own documentation describes it as “a standardized way to connect AI applications to external systems” — the analogy it uses is USB-C: a single physical/logical interface that many devices and many hosts can both target, rather than a proprietary cable per pairing. An MCP server exposes a tool, database, or workflow; an MCP client (built into an agent or IDE) connects to any MCP server without custom glue code for that specific pairing.
It matters that MCP is no longer an Anthropic-only convention. OpenAI documents MCP support in its own API, and MCP clients now ship inside Claude, ChatGPT, Visual Studio Code, and Cursor, among others — the whole point of a shared protocol is that once a server exists for a tool, any MCP-enabled agent can use it, not just the vendor that built the server.
Why it matters when you’re choosing an agent
If you’re evaluating an AI agent to build a real integration on — connect it to your database, your internal APIs, your ticketing system, your CI pipeline — MCP support changes the shape of that work. Without it, “does this agent talk to my systems” is a bespoke engineering question answered by whatever connectors the vendor happened to ship. With it, the agent can reach any existing MCP server (including ones your own team writes once, in-house) without a new custom integration for each new data source. That’s a meaningfully different buying question than “does it have an API” — a documented API tells you the agent can be scripted; MCP support tells you it can plug into a growing shared ecosystem of tools other people have already built servers for.
Our own index, checked: who documents MCP support today
We track 45 researched agent and agentic-tool listings across 7 categories. Rather than guess, we re-scanned every Published listing’s own research write-up — the sourced pros, integrations, and feature copy already gated by our verification process — for an explicit, documented MCP claim, then went back and individually re-checked every listing that wasn’t yet confirmed against its vendor’s current docs, GitHub org or changelog. 35 of 45 (78%) state real MCP support today:
| Tool | Category | What we documented |
|---|---|---|
| Cursor | coding-agents | MCP servers & tools listed as a core integration |
| Claude Code | coding-agents | MCP servers & external tools as a core integration surface |
| Cline | coding-agents | First-class MCP support, with an MCP marketplace for one-click installs |
| Windsurf | coding-agents | MCP servers listed as a core integration, carried over in the Devin Desktop rebrand |
| OpenCode | coding-agents | MCP-server integration layer alongside custom agent skills |
| Bolt.new | coding-agents | Connects out to MCP servers (Notion, Linear, GitHub, custom) — no inbound API |
| OpenAI Codex CLI | coding-agents | MCP servers & tools |
| Devin | coding-agents | MCP Marketplace — client, connects to external MCP servers |
| GitHub Copilot | coding-agents | Copilot Chat and the coding/cloud agent both connect to external MCP servers |
| OpenHands | coding-agents | Client — SSE, Streamable HTTP or Stdio MCP servers via config |
| Replit Agent | coding-agents | Documented as an MCP client in Replit’s own docs |
| Dify | agent-platforms | Publish any app as an MCP-compatible tool |
| Relevance AI | agent-platforms | Agent actions extendable via MCP |
| n8n | agent-platforms | Both directions — an MCP Server Trigger node and an MCP Client node |
| OpenAI Agents SDK | frameworks | MCP servers for external tools, alongside human-in-the-loop approval |
| Pydantic AI | frameworks | MCP servers as a first-class tool source |
| CrewAI | frameworks | Official MCPServerAdapter — client, connects crews to external MCP servers |
| LangChain | frameworks | Official langchain-ai/langchain-mcp-adapters package — client |
| LlamaIndex | frameworks | Both directions — consumes external MCP servers and can publish workflows as one |
| Microsoft AutoGen | frameworks | Official autogen-ext package — client, over STDIO/SSE/Streamable HTTP |
| GPT Researcher | research-agents | MCP server available via a companion project (gptr-mcp), not bundled in the main repo |
| Elicit | research-agents | Official MCP server exposing its search/extraction API |
| Manus | research-agents | Prebuilt MCP connectors (client) to Notion, Linear, Gmail, Stripe and more |
| Perplexity | research-agents | Both directions — an official MCP server, plus local MCP client support |
| Vapi | voice-agents | Custom tools via function calling & MCP |
| ElevenLabs Conversational AI | voice-agents | Both directions — connects to external MCP servers and ships its own official server |
| Retell AI | voice-agents | Both directions — an MCP server for managing agents, plus mid-call client tool calls |
| LiveKit Agents | voice-agents | First-class client support, documented for HTTP and stdio transports |
| Intercom Fin | support-agents | Fully documented API, data connectors and MCP for deeper integration |
| Cresta | support-agents | MCP for standardized tool access, alongside API-based CRM/billing integrations |
| Crescendo | support-agents | Connects its Assistant to external MCP servers, and exposes its own MCP config service |
| Sierra | support-agents | Agents natively publishable to ChatGPT Apps via MCP (server direction) |
| Ada | support-agents | Ships its own MCP server, letting external AI assistants (Claude Desktop, ChatGPT) query Ada’s systems directly |
| Clay | sales-marketing-agents | Runs its own MCP server so Claude, ChatGPT, Copilot and Glean can query a workspace |
| 11x | sales-marketing-agents | “Custom MCP Integration” listed on the Enterprise pricing tier — thin, undetailed |
The category split is the real finding
The first pass of this research made MCP adoption look like a developer-tooling story — coding agents ahead, everything else behind. A full re-check of every unconfirmed listing against primary vendor docs changed that picture: MCP has become close to a market standard everywhere in this index except one category:
| Category | MCP-documented | Share |
|---|---|---|
| Agent frameworks | 6 of 6 | 100% |
| Coding agents | 11 of 14 | 79% |
| Research agents | 4 of 5 | 80% |
| Voice agents | 4 of 5 | 80% |
| Support agents | 5 of 6 | 83% |
| Agent platforms | 3 of 5 | 60% |
| Sales & marketing agents | 2 of 5 | 40% |
Six categories now sit at 60% or higher, five of them at 79–100%. Coding agents and agent frameworks — the two developer-facing categories closest to MCP’s origin as an Anthropic-built, developer-tooling protocol — lead, but the gap to research, voice and support agents (79–83%) is now narrow. Sales and marketing agents remain the clear laggard: only Clay (a full MCP server) and 11x (one undetailed pricing-page bullet) document anything, while Artisan, Qualified and Rox show no MCP mention anywhere in their own materials. That tracks with the category’s usual integration story — purpose-built CRM/outreach connectors sold as a feature, not an open tool-calling protocol the vendor has reason to expose. The honest exception inside a mostly-confirmed field: 11x’s claim is thin enough (a single feature-list bullet, no docs, no stated client/server direction) that we count it as documented but would not call it proven the way n8n’s or Clay’s dedicated MCP docs are — treat that one row as the weakest evidence in this table.
What “not listed” doesn’t mean
The 10 tools not in the table above are not confirmed to support MCP by our research — that is a narrower claim than “these tools don’t support MCP.” Two are genuinely confirmed not yet (Aider has an open, unresolved GitHub feature request for MCP since December 2024; Lindy’s own blog states “it’s not compatible with Model Context Protocol yet”, directly contradicting third-party aggregator claims that it already ships MCP). Two we could not confirm are even the right vendor for a name that collides with unrelated projects (Juggler, Zot). The rest — GC AI, Artisan, Qualified, Rox, Decagon, Bland AI — simply had no MCP mention in their own docs, site or changelog as of this research pass; Decagon and Bland AI are both visibly MCP-aware (Decagon has written blog/glossary content about MCP as a concept; Bland exposes its own documentation site over MCP, though not its voice platform), so a platform-level claim from either would not be a surprise. If you’re evaluating a specific tool not in this table, check its own current docs directly — MCP adoption is moving fast enough in 2026 that this list is a snapshot, not a ceiling.
How to verify MCP support yourself before you build on it
A vendor’s homepage badge is not the same as a working connector. Before you rely on MCP support as a buying decision:
- Look for the vendor’s own MCP server repository (usually on GitHub) or an “MCP” section in their developer docs — not just a marketing mention.
- Check whether the tool ships as an MCP client (it can connect out to other tools’ servers), an MCP server (other agents can connect in to it), or both — they’re different capabilities and vendors don’t always distinguish them clearly.
- Test the connection against a real MCP server before committing an integration to it; “supports MCP” claims vary in how complete the implementation is.
Methodology & sources
The per-tool table above is derived directly from the sourced research already published on each linked listing — not a new survey, and not extrapolated. A tool is counted only when its own listing on this site documents an explicit MCP claim with a cited source; nothing here is inferred from a vendor’s general marketing language. Category percentages are computed from the current Published corpus (45 listings, 7 categories) and will shift as we re-verify existing listings and add new ones — this article is not on an update schedule the way our State of AI Agents 2026 report is, so treat the date above as when it was last checked against the live index.
Primary sources on MCP itself: Anthropic, “Introducing the Model Context Protocol”, 2024-11-25; Model Context Protocol documentation; OpenAI, MCP in the OpenAI API.