UserTesting MCP Servers Put Live User Research Inside AI Chatbots

What happened

UserTesting has released two Model Context Protocol (MCP) servers, one for its core UserTesting platform and one for User Interviews, the recruiting service it bought earlier in the year. Product, design, UX and research teams can now use either one from inside AI applications such as ChatGPT, Claude and Figma Make.1 The Bellevue, Washington company announced the launch on September 24, 2026, through Business Wire. It said the two servers together reach a participant network of more than seven million people across 34 countries.13

The pitch is simple. Teams no longer have to jump between an AI assistant, a research platform and a collaboration tool. They can recruit participants, build a study, launch a test and pull the feedback back without leaving the AI tool where the idea first came up.14 The company says both servers are in Early Access. They support ChatGPT, Claude, Figma Make, Gemini and other MCP-compatible applications, and UserTesting says it will add more capabilities over time.3

The story is bigger than one vendor adding one integration. Two things stand out. First, a research company is treating the AI assistant as a main way into its product, not as an add-on. Second, the launch landed the same week the company changed its corporate name and started selling human judgment as a service for AI products.

What the servers actually do

The announcement lists six capabilities. Teams can recruit participants from within supported AI apps. They can create and launch studies without leaving the AI workflow. Customer insights flow back into the MCP client. AI-generated concepts can be checked with real people before launch. Raw sessions can be turned into summarized findings, with each takeaway tied to the evidence behind it. And teams can move from idea to feedback in one continuous flow, whichever of the two participant networks fits the audience best.14

UserTesting's support documentation shows how this works in practice. Every client connects to a single endpoint, and users sign in with their normal UserTesting credentials. There is no shared or service account, and existing workspace permissions carry over unchanged.9 People with read-only access can connect, find tests and read AI summaries, but creating tests requires write or admin permission.9 One detail stands out: the documentation says test creation is available only through the MCP server. Reading results such as transcripts and quantitative data is already possible through the existing API.9 That is a sign the company sees agents as a first-class way to use the product, not just another copy of what it already offered.

Admins also get controls. In Figma Make, an organization admin publishes the connector once for everyone, and write actions are off by default. If they stay off, nobody in the organization can create or launch a test through Figma Make.9 Accounts that have opted out of generative AI features are excluded from the MCP server entirely.9 On its product page, UserTesting says customer data is not used to train third-party AI models.16

Where the coverage lines up, and where it doesn't

Trade outlets mostly repeated the press release. DestinationCRM, TechIntelPro and Dealroom all described the same workflow, the same list of clients and the same Early Access status.567 Research Live added useful context. It noted that Anthropic developed MCP as an open standard for connecting AI applications and large language models to outside data and tools. It also pointed out that UserTesting is one of several research and insights firms to add MCP support, alongside Stravito and YouGov.8 That puts the launch in context: UserTesting is joining a trend in the insights industry, not starting one.

On the details, reports differ in ways worth flagging:

• Timing. The press release went out on September 24.1 MarTech Cube ran it on September 28, and one analyst wrote that the servers entered early access on September 28, "a day earlier" than the company's September 29 rebrand.2 Those dates don't fit together. The wire release is the more reliable record.
• Release status. The press release says "Early Access."3 A help-center article dated August 14 calls the MCP server an "open beta" that accounts don't need to opt into, yet the same page's FAQ says accounts must be opted in and gives a contact for access requests.9 The marketing page says early access is limited to select teams.16 The safest reading is that access is still controlled, whatever label is used.
• Network size. The launch materials say more than seven million participants.1 The company's July release notes and the rebrand coverage say 7.6 million.17 The two figures likely just reflect different ways of counting the same combined network, not a real conflict.
• Supported clients. The press release names four AI applications.3 The help documentation lists many more, including Claude Code, Codex, GitHub Copilot CLI, Cursor, Visual Studio Code, Amazon Kiro, Gemini CLI and Antigravity IDE.9 So the integration goes well beyond chatbots and into developer tools, which the headline coverage mostly missed.

This has been a while coming

Calling this a "launch" undersells how long it took. UserTesting's April 2026 release page already promoted an early-access MCP server for recruiting participants, launching tests and pulling feedback from Claude, Cursor and other clients.10 At its CRAFTED event, executives said the company had already introduced its MCP capabilities, had a few alpha customers testing them, and had built a second server for User Interviews to create selected audiences.12 The July quarterly release again featured the MCP server, alongside Advanced Targeting across the combined participant network.17

The company's own strategy writing explains why. A March post on responsible AI described a unified architecture built on the Model Context Protocol. It also described "agentic workflows" in which AI might spot patterns, suggest next steps, start a study and draft a report, while humans set the goals and approve the actions.11 Read that way, the September announcement is the moment a long-planned design became available more widely.

Why it matters

The launch materials make one claim that matters more than the feature list. Customer insight now has to be as easy for AI agents to reach as it is for people, because agents are doing more of the work that once meant opening a separate tool.14 That shifts research from a phase in a project to a service that tools can call on demand.

The press release also gives a product-quality argument. AI trained on existing patterns drifts toward the familiar. Products that stand out depend on what real people feel and prefer, which the company says can't be inferred.14 It's a sales message, but it points at something real. As AI tools make generating designs almost free, the scarce thing is a reliable signal about which design actually works for people. Picsart's Oli Mival made the same point in more practical terms: the MCP server keeps real people involved in a much faster development cycle, rather than treating research as a one-time checkpoint.3

Commentary on AI usability tools in 2026 describes the same shift. Agents connected through MCP can set up studies, retrieve data and analyze findings, though human oversight still matters for governance and quality.

The bigger bet behind the launch

My reading is that the MCP servers are less about convenience than about where UserTesting is taking its business. On September 29 the company changed its corporate name to Auros. UserTesting and User Interviews remain as product names, while Auros will sell AI data, evaluation and research services. According to that coverage, chief executive Eric Johnson argued that the hard request isn't a thousand generalists. It's a dozen doctors who can read a medical answer and explain what's wrong with it.

The MCP servers and the rebrand fit together. One lets AI agents order human feedback. The other sets the company up as a source of verified human judgment about AI output. At CRAFTED, the company even demoed a scenario in which panel participants feed preference comparisons and critiques back into model training.12

There are open questions. With Early Access and gated permissions, real-world use is still unproven. And the bigger promise, that insights from different products will eventually merge into one interface, is still described as something that will happen "over time."12 Still, the direction is clear. UserTesting is betting that as AI assistants become the main place people work, the most valuable research platform will be the one an agent can call directly.