Build Log #4: One Person, AI Agents and the Git Log
By MCP Analytics Team | Published , updated
The short version: I built MCP Analytics as one person working with AI coding agents. This post tells that story from the record I can check against: the git history, which keeps a dated entry for every change.
If you want to skip the story and see the product, the case studies are live, or start at the home page. You upload a CSV, a tested statistical method runs on it, and the same data gives the same numbers every time.
Where the history lives
The platform lives in three private git repositories that are still active. A "commit" is one saved change in git, and git records the date of each one. The first commits are dated:
- The MCP server and API: 17 June 2025.
- The module pipeline: 19 Dec 2025.
- Operations (the website, work tracking and scheduled jobs): 25 Dec 2025.
A fourth repository held the first web app. It was shut down on 6 July 2026. An earlier version of this post quoted a commit total; because the repositories are private and nobody outside can check a count, we have taken the numbers out.
Who made those commits
"One person" needs a closer look, because not every commit has a person behind it. Most commits carry my name, from two machines. A large share of the commits in the operations repository were made by scheduled jobs that save their own output (logs, reports, status files); those are automated, not hand-written. A shared account also made commits in the API repository between 16 Sep 2025 and 16 Dec 2025.
Most of the code under my name was written with an AI coding agent (Claude Code) in the session. Those commits carry a "Co-Authored-By: Claude" line, and most commits in each repository have one. That line is a proxy, not a measure of who wrote which lines: it says an agent was in the session, not how much of the change it typed. So the honest summary is: one person directed the work, and AI agents wrote much of it.
The first commit
The first commit is dated 17 June 2025, in the repository that now holds the API. As I remember it, its message was one word: "update." The idea underneath it was to build the analytics as an MCP server. MCP (Model Context Protocol) is a standard way for an AI assistant to call an outside tool. Instead of a separate app that people have to learn, the analysis would be a tool that plugs into the assistant they already use.
The two other repositories started six months later: the module pipeline on 19 Dec 2025 and the operations repository on 25 Dec 2025.
Why build it this way
The starting observation was this: much of the analysis a small business needs is a short list of standard statistical methods applied to its own data. Examples are k-means clustering to group customers, linear regression to see which factors move sales, and cohort retention to see how long signups stay. The answers are specific to each business. The steps to get them repeat.
If the steps repeat, software can run them. So the product takes a CSV and a question, runs a tested method, and returns charts and a plain-English write-up. A person still reads the result and decides what to do with it.
What the agents do
Analysis modules are not written by hand. A pipeline of AI agents, each assigned one stage, plans the approach, writes the code, runs it, checks the output, builds the report and deploys it. My job is to define what a good analysis looks like and to fix what the agents get wrong. The table below lists the parts.
| Part | What it does | First commit |
|---|---|---|
| MCP server and API | Takes the request and the data, returns results | 17 Jun 2025 |
| Module pipeline | Agents plan, code, run, check and deploy an analysis | 19 Dec 2025 |
| Operations | Website, work tracking, scheduled jobs and reports | 25 Dec 2025 |
The operating model
The hardest part of building with AI agents was not writing prompts. It was the coordination layer: how agents hand work to each other, how they flag a problem they cannot solve, and how they keep notes so the next session does not repeat old mistakes. Work items are tracked in a ticket system with numbered entries, and many commits name the ticket they belong to.
An earlier version of this post said the operating model would be released as an open-source framework in June 2026. That release has not happened. As of 2 Nov 2026 there is no public repository for it, and we have removed the promise.
Why one person
This is my experience, not a measured comparison. The common story is that AI replaces workers. What it did here was different: it let me keep track of, maintain and improve more systems at once than I could have alone. On a good day, the scheduled jobs had already written the overnight error report before I started work, so there was less context to rebuild. The git history is the visible trace of that. Not heroics, leverage.
Where to go next
Your turn
Bring your own data and the question you actually need answered.
MCP AnalyticsData science by email Send us your data and your question, and we will reply. ds@mcpanalytics.ai