MCP Servers for Market Research: What Actually Exists in 2026

If you want an AI agent to do market research, you need MCP servers in four categories: survey fielding (asking real people questions), desk research (reading the web), behavioral data (what people actually do), and your own first-party analytics. This post is a working roundup of what exists in each category as of July 2026, with setup notes and the limitations vendors tend to leave out.

Disclosure up front: Chorus Research is our product, and it appears in the survey category. Everything else here is a tool we have either used or evaluated, listed because a roundup that only contains the author's product is an ad, not a roundup.

The quick map

For scale: an analysis of public MCP servers counted over 1,400 public servers by early 2026, and the official MCP Registry passed 9,600 server records by May. Most are developer tools. The research-relevant subset is small, which is why this list fits in one post.

Survey fielding: asking real people

Chorus Research

Chorus is a survey platform built around its MCP server rather than adding one later. An agent can draft a survey, search a qualification catalog to build a target audience (age, region, occupation, product ownership, and so on), validate the spec, get a price, and produce a checkout link. A human pays in the browser, since no tool call can spend money. The agent then polls fielding status and exports structured responses when the quota fills.

• Connection: remote server with OAuth at mcp.chorusresearch.io/mcp, or the @chorus-research/mcp npm package with an API key. Setup steps are in the MCP docs.
• Pricing: per survey, no subscription. Cost scales with audience difficulty and length.
• Limitation: US panel coverage is strongest. Niche B2B audiences can be slow or unfeasible to fill, and the platform will tell you so at estimate time rather than after you pay.

Prolific

Prolific's official MCP server (prolific-oss/prolific-mcp) lets an agent create and manage studies and recruit from a pool of 200,000+ verified participants, with responses returned as structured JSON. It is aimed at human-in-the-loop AI evaluation as much as classic research, and it works well for both.

• Best for: recruiting screened participants for a task or survey you host somewhere else.
• Limitation: Prolific is a recruitment marketplace, not a survey builder. The agent still needs another tool to create the instrument, so the design-field-analyze loop spans two systems.

Qualtrics

Qualtrics ships an official MCP server over its public API with OAuth, covering surveys, responses, and distributions. There is also a more expansive community server (yrvelez/qualtrics-mcp-server, 53 tools) if you want deeper control and accept the maintenance risk of community code.

• Best for: teams already paying for Qualtrics who want agent access to existing studies and response data.
• Limitation: the official server mostly mirrors the existing API surface. Fielding to a paid panel still routes through Qualtrics' separate audience products and pricing, so the agent cannot take a study from idea to fielded data on its own.

SurveyMonkey

SurveyMonkey ships an official MCP server at mcp.surveymonkey.com/mcp with OAuth. It can search, create, edit, and publish surveys, generate shareable collector links, and pull response data and summaries.

• Best for: teams already on SurveyMonkey who want Claude to build and analyze surveys they distribute themselves.
• Limitation: connections are currently limited to Claude, accounts must have US-stored data, and the Audience panel is not exposed, so an agent cannot buy sample or field a study through it. (We compare the platforms more broadly in Chorus vs SurveyMonkey.)

Forms without a panel: Tally

Tally's MCP server lets an agent create and manage forms conversationally. There is no respondent panel (you distribute the link yourself), which makes it a fit for customer-list surveys, waitlist questionnaires, and intercepts where you already own the audience. Community servers exist for Typeform, Jotform, and Google Forms with the same basic shape and the same caveat: the agent can build the instrument but cannot get it answered.

Desk research: Exa and Perplexity Sonar

Before commissioning primary data, a competent researcher checks what is already known. Two servers cover this well:

• Exa does semantic web search built for agents. It returns full parsed content rather than bare links, and can filter to academic papers, news, or company pages. Good default for "what does the published literature say."
• Perplexity Sonar exposes Perplexity's answer engine as a tool call: synthesized answers with citations. Faster than Exa when you want a conclusion rather than source documents, weaker when you need to verify the underlying material yourself.

Both are API-key setups with usage-based pricing. The failure mode to watch: agents treat confident secondary summaries as settled fact. For anything decision-grade, use desk research to sharpen the question, then test it with primary data.

Web data at scale: Firecrawl and Bright Data

When the research question is "what do competitor pricing pages say" or "collect every review of this product category," you need scraping rather than search. Firecrawl converts sites into clean, LLM-ready markdown and handles crawling logic; it is the simpler of the two. Bright Data's MCP server is the heavyweight option: proxy infrastructure that gets through anti-bot measures, with pricing and complexity to match. Start with Firecrawl and graduate to Bright Data when you get blocked. Mind the legal side: scraping public pages is generally defensible, but respect robots.txt and site terms, and never collect personal data this way.

Behavioral and trend data

Similarweb

Similarweb publishes an official MCP server over its digital-intelligence dataset: traffic estimates, engagement metrics, and audience overlap for essentially any significant website. For competitive sizing questions ("is their product growing?") it replaces hours of manual lookup. It is enterprise priced, and traffic estimates for smaller sites carry wide error bars, so treat the numbers as directional rather than exact.

Trend servers

Trends MCP aggregates live trend data across Google Search, YouTube, TikTok, Reddit, Amazon, and app stores, with a free tier. Several community Google Trends servers (for example andrewlwn77/google-trends-mcp) cover keyword interest, related queries, and regional breakdowns. These are useful for timing and demand questions, though they tell you what people search for and nothing about the reasons behind it.

PostHog

PostHog's official MCP server gives agents access to your own product analytics: events, funnels, feature flags, session data. In a research stack it answers the "what are users actually doing" half of the question, so your surveys can focus on why. Amplitude ships an official MCP server as well (OAuth, respects existing permissions); GA4 access still runs through community servers, so check maintenance activity before trusting one.

How to assemble a stack

A useful minimal setup for a product team is three servers:

1. One desk-research server (Exa or Perplexity) to establish what is known.
2. Your first-party analytics server (PostHog) to see current behavior.
3. One survey server with real respondents (Chorus, or Prolific if you host studies elsewhere) to test what you cannot observe.

That covers secondary, behavioral, and primary data with three API keys. Add scraping and trend servers when a specific project demands them. Connections you rarely use add attack surface without adding much capability.

On that note, two security habits matter. First, prefer OAuth-based remote servers or scoped API keys over full-access credentials. An agent holding your admin key can do anything the key can. Second, read what tools a server exposes before connecting it. Enterprise adoption surveys show MCP moving into production (41% of surveyed software orgs), which is all the more reason to be careful about what you wire into your agent.

What is still missing

Gaps you should know about before you go looking: there is no good MCP server for qualitative research (moderated interviews and focus groups), the enterprise panel marketplaces Cint and Dynata ship no official servers, and conjoint or MaxDiff studies remain out of agent reach on every platform including ours. If your research program leans on those methods, agents can currently help with the work around them but not the methods themselves.

For background on what changes when research becomes something agents do, rather than something people do with software, see our explainer on agentic market research.