Research report · ai-agents
What Is MCP, and Which AI Agents Actually Support It?
Of the 94 AI agents in our index on 27 August 2026, 63 document MCP support. Sales and marketing agents remain the laggard, now two of eight.
On this page
Documented MCP support by category
Of the 94 AI agent product listings in the index (documented MCP support is a per-listing claim, checked against source), the largest group is “Coding agents” at 19 (20%). Measured 2026-08-27.
Of the 94 AI agents published in our index on 27 August 2026, 63 document support for the Model Context Protocol. That is a count of what these tools say in their own docs, repositories and changelogs. It is not a survey of the market, and the method section below states exactly what we counted and what we did not.
Almost every listing page in this space now advertises MCP support somewhere, usually with no explanation of what it buys a buyer. Below is the plain definition, the reason it matters when you are choosing an agent to build on, and the count from our own index.
What MCP actually is
The Model Context Protocol is an open standard, introduced by Anthropic on 2024-11-25 and open-sourced the same day, for connecting AI applications to external data 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.” That is the N-times-M problem, where N AI applications each need a bespoke connector to M tools. MCP replaces the grid with one protocol both sides implement once.
The protocol’s documentation describes it as “a standardized way to connect AI applications to external systems,” and reaches for USB-C as the analogy: one 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 connects to any MCP server without glue code written for that specific pairing.
MCP is no longer an Anthropic convention. OpenAI documents MCP support in its own API, and MCP clients now ship inside Claude, ChatGPT, Visual Studio Code and Cursor. Once a server exists for a tool, any MCP-enabled agent can reach it.
Why MCP support changes what an integration costs you
If you are evaluating an agent to connect to your database, your internal APIs, your ticketing system or 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 per data source. That is a different buying question from “does it have an API.” A documented API tells you the agent can be scripted. MCP support tells you it plugs into a shared ecosystem of servers other people have already built.
Our index, counted on 27 August 2026
Of the 94 AI agent products published in The Agents Index on 27 August 2026, 63 document support for MCP as a client, a server, or both. The remaining 31 are covered in their own section below, and 31 is not a count of tools that lack MCP. The corpus also holds one agency listing, Agentix Labs, which sells engineering services rather than a product with an integration surface of its own, so it sits outside this count entirely.
Read 63 as a floor rather than a market share. We do not survey vendors. A tool is counted only where its own listing on this site already carries an explicit MCP claim with a cited source, drawn from the sourced integrations, feature copy and pros our verification process has gated. A tool that supports MCP but has not yet had that fact researched onto its listing counts against us, not for us. The full current membership is maintained as a standing set: AI Agent Tools With MCP Support.
Sales and marketing agents narrows the gap, but stays last
Seven of the eight categories in our index document MCP on at least half their members. Sales and marketing agents document it on two listings in eight — the lowest share in the index, but no longer the one-in-seven outlier it was two days ago.
| Category | Documenting MCP | In the index | Share |
|---|---|---|---|
| Agent frameworks | 12 | 14 | 86% |
| Coding agents | 19 | 25 | 76% |
| Research and data agents | 8 | 11 | 73% |
| Agent platforms | 11 | 16 | 69% |
| Support agents | 6 | 9 | 67% |
| Voice agents | 7 | 12 | 58% |
| Agent tools | 1 | 2 | 50% |
| Sales and marketing agents | 2 | 8 | 25% |
(Botsify, Sierra and ElevenLabs Conversational AI each belong to two categories and are counted once in each. All three document MCP, so these rows sum to 97 members and 66 documenting against 94 distinct tools and 63 distinct documenting. The same convention applies in our deployment census.)
11x and the newly-added Lead Scorer are now the only sales and marketing agents in our index documenting MCP. Clay, the category’s other documented listing two editions ago, left the index before this edition. The category has grown from 5 listings in July 2026 to 8 today, and the count of listings documenting MCP went from two down to one, and is now back up to two — a recovery driven entirely by the new listing rather than by an existing vendor shipping support it lacked before. The likely reason the rest of the category still lags is commercial rather than technical: most of these tools sell purpose-built CRM and outreach connectors as the product, and an open tool-calling protocol makes those connectors easier to route around.
11x’s own evidence is still weak, and we’d rather say so: it lists “Custom MCP Integration” as a bullet on an enterprise pricing tier, with no docs and no stated client or server direction. It counts as documented under our rule, but it remains the thinnest entry in the whole set. Lead Scorer sits at the opposite end: MCP isn’t a bullet point but the product’s entire delivery mechanism, a single OAuth 2.1 server exposing 112 tools, documented on its own MCP page. The category’s documented pair now spans the full range of what “documents MCP” can mean on this site.
What “not documented” does not mean
Thirty-one of the 94 product listings in the index are not confirmed by our research to support MCP, which is a narrower claim than “these 31 tools do not support MCP.” The difference is the whole point of this section, and it is where a census like this one is easiest to misread.
Thirteen of the 31 are documented absences with a specific reason, taken from each listing’s own current text: Aider, Lindy, GC AI, Artisan, Qualified, Rox, Decagon, Bland AI, Juggler, Zot, Unify, Moveworks and Vecbase. Moveworks describes MCP as a capability it is developing, framed on its own blog as planned routing of lower-priority tasks rather than a shipped feature. Aider has an open, unresolved GitHub request for it since December 2024.
AgentsKit is an absence of a third kind: it exposes an /api/mcp discovery endpoint that indexes its own catalogue, which is neither an MCP client nor an MCP server.
Fourteen more carry no MCP claim in our research at all, and have not been read for it since they were added: Amplemarket, Auto-Respond, Browserbase, Deepgram Voice Agent API, Genspark, Julius AI, Junie, LangGraph, Muse Code, OpenClaw, SkipCalls, Synthflow AI, Warmly and Zendesk AI Agents.
The remaining three are simply too new to have been triaged either way: Vellum and VoiceAgent joined the last edition, and Kilo Code joined the one before that. Nothing has been established about any of the 31 in either direction, and we would rather record that than record a No we never verified. If you are evaluating a specific tool that is not in the documenting set, read its current docs directly. Adoption is moving fast enough in 2026 that any count of this kind is a snapshot rather than a ceiling.
What the July 2026 pass measured, and why it is not a trend line
We first ran this count on 25 July 2026, against a 45-listing index, and recorded 35 of 45 documenting MCP. Do not read 35 of 45 against 63 of 94 as a change in adoption. The two figures were produced by different methods and are not comparable.
The July figure came from an active pass that took every unconfirmed listing and re-checked it against the vendor’s current docs, GitHub organisation or changelog. The current figure is a passive count of what our listings already document, across an index that has since roughly doubled. The gap between the two is a statement about our research coverage, not about the market.
That July pass also recorded individual not-yet verdicts on ten named tools. We are not carrying those forward here, because a negative about a named vendor goes stale faster than a positive: shipping MCP is an announcement, and an absence is only ever an absence on the day somebody looked.
How to check MCP support yourself before you build on it
A vendor’s homepage badge is not 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. A marketing mention is not either of those.
- Check whether the tool ships as an MCP client, which connects out to other tools’ servers, or an MCP server, which other agents connect into, or both. They are different capabilities and vendors do not reliably distinguish them.
- Test the connection against a real MCP server before you commit an integration. Implementations vary in how complete they are.
Method and sources
The count is derived from the sourced research already published on each listing in our index, not from a new survey and not extrapolated. A tool is counted only when its own listing documents an explicit MCP claim with a cited source, as a client or a server. Nothing is inferred from a vendor’s general marketing language, and a machine-readable discovery endpoint does not qualify as either side of the protocol.
The figures above were computed from the Published corpus of The Agents Index as it stood on 27 August 2026: 94 product listings across 8 categories, of which 63 document MCP support and 31 do not, plus one agency listing held outside this count. Since the last edition, Lead Scorer joined and ships as an MCP server itself, so it is counted immediately; Clay had already left the index the edition before, dropping its category’s documented count from two to one, and Lead Scorer’s arrival brings that category back to two. Vellum, VoiceAgent, Agentix Labs and Kilo Code remain unreviewed for MCP. The July 2026 figures are frozen at the 45-listing corpus of 25 July 2026 and were produced by a different method. All of these numbers shift as we re-verify existing listings and add new ones, so treat the date above as the day this 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.
Get the next report
New agents rankings and fresh data reports. One short email, one-click unsubscribe.