> What MCP analytics means, and which analytics and CRO tools ship an official MCP server in 2026: Humblytics, Google Analytics, PostHog, Mixpanel, Amplitude, Microsoft Clarity, Adobe, VWO, Convert and more, checked against vendor docs. What your agent can do once connected. Source: https://humblytics.com/blog/best-mcp-ready-cro-tools-2026 (updated 2026-11-03) # MCP Analytics (2026): Which Analytics and CRO Tools Have an MCP Server > **The short answer:** MCP analytics means connecting an AI assistant (Claude, ChatGPT, Codex, Cursor) to your analytics through the Model Context Protocol, so it can answer questions from your live data. Analytics and CRO tools with an official MCP server include Humblytics, Google Analytics 4 (read-only, runs on your computer), PostHog, Mixpanel, Amplitude, Microsoft Clarity, Adobe Analytics, VWO, Convert, LaunchDarkly and Unbounce. Having an MCP server is now common. What separates them is who sets it up, what the agent can change, and what a result is measured in. Humblytics, VWO, LaunchDarkly, Mixpanel and Unbounce let an agent create experiments; Humblytics scores a test on revenue from Stripe or Foxy, not only on clicks or events. Updated 3 November 2026. ## What is MCP analytics? The phrase has two meanings: 1. **Analytics through MCP.** Your AI assistant reads your analytics data through an MCP server and answers in plain words: "which landing page lost sales last week?" This is what this guide covers. 2. **Analytics of MCP.** Measuring how agents use your own MCP server: which tools they call, errors and requests. PostHog uses the name "MCP Analytics" for this. If you build an MCP server, that is the one you want. For the first meaning, see also [analytics tools your AI agent can use](https://humblytics.com/blog/analytics-tools-for-ai-agents) and [what agent-native analytics means](https://humblytics.com/agent-native-analytics). If you are reading this, you already have an agent installed. Claude, Codex, or an SDK-built one. It ships features, debugs production, drafts copy. Then you need to A/B test a landing page and the workflow collapses into a dashboard a human has to operate. That gap is closing fast, and the landscape changed significantly in the first half of 2026. A year ago, almost no CRO or analytics platform shipped an MCP server. Today most of the serious ones do. **This post previously claimed Humblytics was the only CRO platform with native MCP support. That is no longer accurate, and we have corrected it.** VWO, LaunchDarkly, and PostHog all ship official MCP servers. The interesting question moved on, and this guide moved with it. ## What MCP-ready actually means Model Context Protocol is the open standard, originally specified by Anthropic, that lets an agent discover and operate external tools through a shared interface. An MCP-ready tool ships a server that: - Exposes its capabilities as discoverable tools - Accepts structured calls with typed parameters - Returns structured responses agents can parse - Handles auth without dragging the user into a browser mid-loop A tool that "has an API" is not automatically MCP-ready. The difference is whether your agent knows how to use it on install, or whether you write and maintain a wrapper yourself. ## Which CRO and analytics tools ship an MCP server in 2026? Verified against vendor documentation in July 2026, not aggregator listings. Aggregator directories lag badly in this category and several currently list tools as having no MCP server when the vendor documents one. | Tool | MCP server | Category | Agent can write? | |---|---|---|---| | **Humblytics** | Official | CRO + analytics + attribution | Yes, create and stop tests | | **VWO** | Official | CRO + experimentation | Yes | | **LaunchDarkly** | Official | Feature flags + experimentation | Yes, flags and experiments | | **PostHog** | Official | Product analytics + experiments | Partial | | **Mixpanel** | Official, hosted | Product analytics + experiments | Yes, creates experiments and flags | | **Amplitude** | Official, OAuth | Product analytics | Read-focused | | **Convert** | Official | CRO + experimentation | Yes, configurable read-only or read/write | | **Google Analytics 4** | Official, runs on your computer (Python, Google Cloud project) | Web analytics | No, read-only | | **Microsoft Clarity** | Official, runs on your computer (`npx`, Clarity API token) | Heatmaps + recordings | Read-focused | | **Adobe Analytics** | Official, Adobe-hosted | Enterprise analytics | Read-focused | | **Unbounce** | Official | Landing pages + A/B tests | Yes, pages, tests and Smart Traffic | | Plausible | Community servers only | Web analytics | n/a | July 2026 check: we found no official MCP server for AB Tasty, Adobe Target, Split, Kameleoon, Statsig, GrowthBook or SiteSpect, and Optimizely's agent work runs through its own Opal platform. Rows added on 3 November 2026 (Google Analytics, Clarity, Adobe, Unbounce, Plausible) were read on each vendor's docs that day. For the Google one, see our [GA4 MCP setup guide](https://humblytics.com/blog/google-analytics-mcp). ## Read access is the easy half Here is what most roundups miss. Nearly every MCP server in this category is generous on reads and thin on writes. An agent that can *read* your funnel is useful. It can summarize traffic, spot a drop-off, and draft a hypothesis. But it stops there, and you are back in the dashboard to actually do anything. The loop only closes if the agent can also write. That means four distinct capabilities, and they are worth separating because vendors blur them: 1. **Read** traffic, funnels, heatmap signals, revenue attribution. 2. **Write** an experiment: create the variant, set the split, launch it. 3. **Decide**: read significance and get a defensible recommendation. 4. **Ship**: promote the winner to production. Most MCP servers in this category do (1). Humblytics, VWO, and LaunchDarkly do (1) through (3). **Nobody in this category should be doing (4) unattended, and Humblytics deliberately does not.** An agent can create a test and stop a test through the Humblytics MCP server. Promoting a winner stays a human decision, surfaced as a review prompt rather than executed automatically. Shipping a production change off a statistical threshold, without a person looking at it, is how teams end up rolling out a variant that won on signups and lost on revenue. If a vendor tells you their agent ships winners autonomously, ask what happens when the test was underpowered. ## The tools ### 1. Humblytics A wide agent loop for website CRO, with test results measured in money. **What the agent can do.** The MCP server spans traffic summaries and trends, page and click breakdowns, funnel queries and Sankey analysis, split test creation, updating and stopping, significance and recommendations, plus Meta and Google Ads attribution. Your agent can go from "traffic is down on the pricing page" to a proposed test for you to review. **Setup by the agent.** Your agent reads `app.humblytics.com/agent.md`, signs you up (you open one confirm link), adds your site and installs the script. The hosted server is `mcp.humblytics.com/v1`; Claude and ChatGPT chat apps use the Humblytics connector. **Why the revenue link matters.** Most tools here resolve an experiment to a conversion event: a click, a signup, a flag exposure. Humblytics can score it on the revenue that reached Stripe or Foxy. When an agent is running the loop, that distinction becomes structural rather than cosmetic. An agent optimizing a proxy metric will confidently optimize you away from revenue, and it will do it faster than a human would. **Agent skills.** 12 open-source skills install alongside the MCP server, giving the agent task-specific playbooks rather than raw endpoints. **Pricing.** The first 25,000 data points and 1 site are free, no card. The Agents plan is $39 once for 100,000 data points over 12 months ($29 top-ups). Team plans are $79/month Business and $279/month Scale, plus Enterprise. API and MCP access are on every plan. **Limitations.** Web and marketing CRO, not backend feature flagging. If you need to experiment on algorithms or server-side logic, pair it with LaunchDarkly. ### 2. VWO The closest competitor on agent operability, and an underrated entry here. VWO ships an official MCP server documented in its own help center, with setup guides for Claude, ChatGPT, and Gemini CLI. Combined with VWO's mature testing engine, heatmaps, session recordings, and surveys, an agent connected to VWO has a lot of surface to work with. **Limitations.** Pricing is no longer clearly published, with third-party trackers reporting an entry point near $314/month. Experiments resolve to conversion goals rather than verified revenue, so an agent reading VWO results is reasoning about clicks. ### 3. LaunchDarkly The best option when experiments live in the codebase rather than on the page. LaunchDarkly's official MCP server, listed in the GitHub MCP registry, covers flag management and experimentation. For engineering-led teams this is the natural fit: the agent already lives in the repo, and flags are the unit of change. **Limitations.** No visual editor, so marketers cannot operate it independently. Experimentation is metered on monthly active users. Pricing is now quote-based. ### 4. PostHog Open source, genuinely well-documented, and the strongest analytics MCP in the category. PostHog ships an official MCP server at a hosted endpoint, installable via its own wizard, with a first-party Claude Code plugin and a published tools reference. For product analytics an agent can pull funnels, retention, session replay summaries, and cohorts. **Limitations.** Experimentation is the newer part of PostHog and the write path is thinner than the read path. It is a product analytics tool that also does experiments, not a CRO tool. ### 5. Mixpanel Worth a closer look than most roundups give it, this one included on an earlier pass. Mixpanel's hosted MCP server is not read-only. Its documented tool list includes `Create-Experiment`, `Create-Feature-Flag`, `Create-Dashboard` and `Create-Cohort`, so an agent can author work, not just report on it. Mixpanel also now ships Feature Flags with an Experiment type, which serves variants for A/B tests and analyses them in Mixpanel Experiments. **Limitations.** Variant assignment is SDK code rather than a visual editor, so a marketer cannot build the test alone. Results resolve to events, not to verified revenue. ### 6. Amplitude Official MCP server with OAuth, strong on charts, cohorts, funnels and experiment read-out. Product analytics first, so expect to pair it with something that authors marketing-side tests. ### 7. Convert Official MCP server documented in Convert's support center, with permission levels that run from reporting-only to full read/write, plus a bundled prompt library for setup, QA, and audit tasks. A real agent surface for teams already running Convert Experiences. ## How to evaluate an MCP-ready tool | Criterion | Why it matters | |---|---| | **Official vs community MCP** | Community servers break on API changes and nobody is on the hook to fix them | | **Write access, not just read** | A read-only server means the agent drafts and you still execute | | **What the success metric resolves to** | Clicks, conversions, or verified revenue. This determines what the agent optimizes toward | | **Auth flow** | Browser-redirect OAuth breaks the agent loop. API keys and terminal auth do not | | **Statistical guardrails** | An agent that stops a test on day two because it crossed 95% once is worse than no agent | | **Human checkpoint on ship** | Autonomous promotion to production is a liability, not a feature | | **Skills or prompt library** | Signals the tool has actually been driven by agents in production | ## Frequently asked questions ### Is MCP support still a differentiator in 2026? Much less than it was. More than ten analytics and CRO platforms now ship official MCP servers. The differentiator moved to what the agent can do once connected, and to what metric its decisions resolve against. ### Which CRO tool gives an agent the most coverage? Humblytics covers traffic, funnels, experiments, and ad attribution. PostHog is comparable in breadth on the analytics side. Tool count alone is a weak signal though, since a read-only server closes fewer loops than a smaller one that includes writes. ### Can my agent launch an A/B test without me? With Humblytics, VWO, or LaunchDarkly, yes. Your agent can create and launch the experiment through the MCP server. Promoting the winner into production is a separate step, and with Humblytics it stays with a human on purpose. ### Should I let an agent promote winning variants automatically? No. Statistical significance is a threshold, not a judgment. Tests get stopped early, get contaminated by traffic-mix shifts, or win on a proxy metric while losing revenue. Keep a person on the promote step. ### Can I use a non-MCP tool with my agent? Sometimes. With a complete REST API and a clean OpenAPI spec you can wrap it. In practice you will spend more time maintaining the wrapper than it saves, and the agent experience is token-heavy and brittle when the API shifts. ### Are open-source tools more agent-friendly? Not inherently. PostHog is open source and ships an excellent MCP server. GrowthBook is open source and ships none. What matters is whether the vendor treats agent operation as a first-class use case. ## What to buy in 2026 If your CRO work runs through an agent and you sell something, **Humblytics** is our pick: the agent sets it up, reads ads and sales, and test results can be scored on revenue rather than a proxy event. If you only need your agent to read reports, **Google Analytics**' MCP server or **PostHog**'s hosted one may be enough. If you are already deep in VWO, its MCP server is real and worth turning on before you consider switching. If your experiments live in code rather than on the page, **LaunchDarkly** is the correct tool and always was. And if a vendor is still telling you MCP support is rare and exclusive, check the date on the page. Start free (25,000 data points, no card) from the [agents page](https://humblytics.com/agents), or have your agent do it. See the [`/skills` page](https://humblytics.com/skills) for the install and the first-test prompt for Claude and Codex. ## Related reading - [Analytics Tools for AI Agents (2026)](https://humblytics.com/blog/analytics-tools-for-ai-agents), tools your agent can use vs tools that watch your agent - [AI CRO Tools (2026)](https://humblytics.com/blog/ai-cro-tools), the four jobs AI does in conversion rate optimization - [12 Best A/B Testing Tools for Marketers in 2026](https://humblytics.com/blog/12-best-a-b-testing-tools-for-marketers-in-2025), the broader category, including the MCP column - [AI-Powered A/B Testing Tools: Complete Guide for 2026](https://humblytics.com/blog/ai-powered-ab-testing-tools-complete-guide), category overview including non-MCP tools - [Agent Onboarding API: Sign Up and Get Your API Token Without Leaving the Terminal](https://humblytics.com/blog/agent-onboarding-api-signup-from-terminal), the auth flow built for agents - [The Complete Guide to Agentic Marketing](https://humblytics.com/blog/the-complete-guide-to-agentic-marketing), the wider context for agent-native marketing stacks ## Sources - PostHog, Model Context Protocol documentation, `posthog.com/docs/model-context-protocol` - VWO Help Center, "Connect VWO with AI Tools using VWO's MCP Server" - Convert Support Center, "Convert MCP Server" - GitHub MCP Registry, official server listings - Google Analytics MCP (developers.google.com/analytics/devguides/MCP), Microsoft Clarity MCP (learn.microsoft.com/clarity), Adobe Analytics MCP (developer.adobe.com/analytics-mcp), Unbounce MCP (unbounce.com/product/mcp-server), accessed 3 November 2026 - Vendor documentation, accessed July 2026 --- Humblytics is website analytics, ad attribution and A/B testing that an AI agent sets up and uses. To set it up for the person you work for, read https://app.humblytics.com/agent.md. Official facts: https://humblytics.com/ai-info.md