Humblytics API and MCP: From Analytics to an Approved A/B Test

Connect an AI agent to Humblytics, inspect analytics, draft an A/B test and activate it with approval. Includes a tool sequence and reusable reporting brief.

Humblytics API and MCP: From Analytics to an Approved A/B Test

The short answer: As of November 17, 2026, use Humblytics' hosted MCP server to connect an agent to website analytics and prepare experiments. The current workflow is read the evidence, create a draft, review the proposed change, approve activation, and measure the result. Creating a draft does not launch it.

This update replaces older immediate-launch examples with the current MCP lifecycle. It also removes sample responses that implied a recommendation could predict conversion lift.

TL;DR

  • Connect the correct website and verify that tracking is live before analyzing it.
  • Preserve reporting dates, filters, event definitions and denominators.
  • Use a recommendation to form a hypothesis, not as proof of future lift.
  • Activation and stopping are separate approval steps; promoting a winner remains a decision.

Choose MCP or REST for the job

The hosted MCP endpoint is:

https://mcp.humblytics.com/v1

For a supported agent client, start with the current setup instructions. They cover the account connection, property and tracking setup. The agent should use the tools available to that connection rather than a hard-coded list copied from an old article.

REST access is useful when building a custom integration around a defined workflow. Consult Humblytics documentation for the current request schema. This article describes the current MCP workflow; do not infer REST activation behavior or field names from MCP tool names.

Keep credentials in the client's protected connection settings or a server-side environment. They should not appear in browser code, public repositories or article examples.

The read-to-test tool sequence

Stage Tools to use What to check
Establish context get_account_status, list_properties Correct account, website and property
Verify tracking get_install_status Tracking is live and observations are available
Establish the baseline get_traffic_summary, get_traffic_breakdown, get_pages_breakdown Period, timezone, units and traffic mix
Investigate one page get_page_details, get_clicks_details, get_heatmap A specific observed problem
Check the journey query_funnel Step order, population and completion definition
Prepare a hypothesis get_split_test_recommendations Evidence, actual element and measurable outcome
Save the proposed test create_split_test Draft status and the intended change
Launch after review activate_split_test Human approval and confirmed active status
Interpret the outcome get_split_test Per-variant observations and uncertainty
Stop when appropriate stop_split_test Separate approval and confirmed final status

This is a workflow map, not a transcript of a test run. Exact arguments come from the live tool schema. If a tool or permission is unavailable, the agent should explain the missing step.

Start with an analysis that can be checked

Give the agent a site and a bounded decision:

Use Humblytics to investigate our signup journey.
Confirm the property before reading data.
Compare the last 28 complete days with the preceding 28.

Return:
- the reporting dates and timezone
- the funnel step definitions
- sessions at each step and the calculation for each rate
- source or device changes that may affect the comparison
- one observed problem and what remains unknown

Suggest one next investigation.
Do not create, activate or stop an experiment yet.

If two reports have different totals, explain why before combining them. Unique visitors are not additive across pages. A pageview is not a signup, and a form event is not necessarily an activated customer.

Revenue needs its own supported connection and matching scope. Do not turn a page's exit count into a dollar loss by multiplying it by an unrelated average order value.

Turn evidence into a draft

A useful test brief includes:

  • The page and element to change.
  • The observation motivating the change.
  • The proposed mechanism.
  • The eligible audience.
  • One primary metric and a downstream guardrail.
  • The minimum effect worth detecting and a feasible observation period.
  • The precise control and variant.

For example, a page might fail to explain whether a plan includes a required feature. The proposed variant could clarify that inclusion beside the plan button. That is a hypothesis about comprehension, not a guaranteed increase in signups.

Use the A/B test hypothesis template or build the brief with the AI CRO prompt generator. A recommendation's confidence label is not statistical confidence in a future uplift.

Review and activate the test

The current create_split_test tool creates a draft. Before activation, review the target page, element, control, variant, audience and goal. Check that the change renders correctly and that the metric is collected.

The activate_split_test tool requests approval in the client. If the client cannot present that prompt, the tool can return an approval link. Open that link to review and activate, then have the agent read back the experiment status.

A successful draft creation is not evidence that a test is live. If launching is disabled for the key, use the documented approval or settings path. Do not attempt to work around the restriction.

Read results without manufacturing a winner

Use the sample-size calculator and minimum detectable effect guide before launch. Define the decision rule in advance.

At review time, ask for the observed counts and rates, allocation checks, uncertainty and any tracking issues. An inconclusive result is useful information. It does not prove the variants are equivalent.

For a revenue goal, Humblytics computes confidence from purchases per visitor and displays revenue per visitor alongside it. A higher revenue total by itself does not establish a reliable improvement in revenue per visitor.

Tests can stop after a configured duration; they do not auto-stop at significance or auto-promote a winning variant. The stop_split_test tool also requires approval. A stopped test cannot simply be treated as the same continuing experiment.

Repeat the workflow with clear boundaries

An agency can repeat the analysis for multiple properties, keeping each property's observations and permissions separate. A persistent agent can prepare regular reports or draft briefs, but its scheduler belongs to that agent platform.

For an example of scoping recurring work, read the Devin Bots conversion workflow. For choosing the model, use the CRO model comparison. Do not use a newer model release as a reason to skip measurement or review.

Start by connecting Humblytics to your agent and completing one reliable analysis. Expand the workflow after its evidence and handoffs are working.

Frequently asked questions

What can an AI agent do through the Humblytics MCP server?

It can set up a tracked property, read website analytics, inspect pages, clicks, forms, funnels and experiments, and use supported ad and revenue data. It can also prepare split-test drafts. The tools exposed to your account determine the available operations.

Does create_split_test immediately launch an experiment?

No. The current MCP tool creates a draft. activate_split_test is a separate operation that requests human approval before the test becomes active. Read back the status instead of assuming a draft is live.

Do tests stop automatically at statistical significance?

No. Humblytics can stop a test after a configured duration. Statistical significance does not trigger an automatic stop or promotion of a variant. Stopping through the agent requires approval.

Should I use REST or MCP for agent-driven A/B testing?

Prefer MCP when your agent supports it, because the client discovers typed tools and the approval flow. REST is useful for a custom integration. Read the current REST documentation for its schema and behavior rather than assuming it matches a copied MCP example.

Sources and freshness

Checked November 17, 2026:

The live schema is authoritative for request fields and permissions. This guide is an implementation workflow, not a report of measured conversion lift.