Build Log #1: The Starting Line

This is the first entry in our build log, a public record of how MCP Analytics is being built. Each entry says where we are, what is working, what is not, and the numbers behind it. Where a number appears, the page says where it came from.

What MCP Analytics is

MCP Analytics is a statistical analysis and reporting product. You upload a CSV (a plain table of rows and columns, such as a Shopify export or a marketing spreadsheet). Tested statistical methods run on it, and you get back a report: charts, statistical tests, checks on the method's assumptions, a plain-English summary, and the code that produced it. It is repeatable: the same data gives the same numbers.

The methods are written in R, a free language for statistical computing. In April 2026 there were two ways in: a web app, and an MCP server. MCP (Model Context Protocol) is an open standard that lets AI apps such as Claude connect to outside tools, so a developer could run an analysis from their AI assistant.

Update: the original web app was shut down on July 6, 2026. The customer site is now account.mcpanalytics.ai.

How it was built: the commit history

I built this mostly by writing instructions for AI coding agents (Claude Code) and reviewing what they produced. That is my own account, so here is a record anyone with the code can check. Every commit (a saved change to the code) that an agent helped write carries a line reading "Co-Authored-By: Claude". This counts all commits, and those with that line, in three code repositories up to the day this post went out:

git rev-list --count --before='2026-04-08T00:00:00' HEAD
git rev-list --count --before='2026-04-08T00:00:00' -i --grep='Co-Authored-By: Claude' HEAD

Across the three repositories, about 66% of commits made up to April 7, 2026 carry that line; in each repository on its own the share ranges from about 37% to about 94%. Counts were taken on November 2, 2026. Read them with care. The co-author line is a proxy, not a measure of who wrote the code. It is also a lower bound on agent help: some agent-written commits, such as automatic commits made before a deploy, do not carry it. A commit count also says nothing about how large or how good each change was. What the counts do show is that most saved changes in this period went through an AI agent, not that the agents worked without a person deciding what to build and checking it.

In practice the analysis modules (each one an R script for one statistical method) start from a written specification. The agents write the R code and run its tests, and I decide what ships. Day-to-day jobs such as reading error logs, pulling traffic numbers and deploying code are run as saved Claude Code commands that I start by hand.

Search traffic before this post

Here is what Google Search Console reports for web search over the 30 days before this post, March 8 to April 6, 2026. An impression means a link to our site was shown in a Google result; a click means someone followed it.

That is 1,109 clicks out of 353,064 impressions, about 0.31%. Source: Search Console property sc-domain:mcpanalytics.ai, search type web, grouped by device, queried November 2, 2026. Google can revise past data, so a later query may differ slightly.

We were seen often and clicked rarely. My reading at the time, which is a judgment and not something this data proves, was that our method articles (explainers such as "what is logistic regression") reached people studying statistics more than people with a business question to answer.

The plan in April 2026

These were plans when this post was written, not results:

• Business-question content. Write more pages around questions a business asks, such as "how do I analyze my Shopify churn", alongside the method explainers.
• A faster first report. Give new users sample datasets so they can see a full report before uploading their own data.
• In-report feedback. Let readers rate each slide of a report and flag problems, and send repeated flags into our fix queue for a person to review.
• Use it on ourselves. Run our own business data through the product.
• Share the tooling. Open-source parts of the task-tracking system we use to run the work. As of this update that has not happened.

Later entries in this log report how each of these went.

Why this log exists

Building a software business where AI agents write most of the code is still new, and there are few written records of what it looks like day to day. This log is ours: what works, what does not, and the numbers, with their sources.

Keep reading

• Next: Build Log #2, What AI Actually Does for a Founder
• All build log entries