DEV Community

kyb8801
kyb8801

Posted on

What I learned shipping an MCP server alongside an Excel toolkit

#metrology #iso #mcp #productivity

Last week I shipped two things in 72 hours:

  1. A free MCP server (50 calls/month, hosted) that exposes 10 typed tools for measurement uncertainty math — Type A/B, Welch-Satterthwaite, Monte Carlo, the lot.
  2. A paid Excel toolkit ($19) that wraps the same math in a shape an ISO/IEC 17025 calibration lab can actually drop into Monday morning.

The shape of that split is the post. Not the math, not the niche — the
operating model.

Why split free MCP + paid Excel?

The MCP server is the discovery surface. Engineers add it to Claude
Desktop in one CLI command. They ask "compute Type A from these 10
replicate measurements" and the LLM picks the right primitive and
returns a typed JSON answer. Cost to the user: nothing.

The Excel toolkit is the production surface. ISO/IEC 17025 auditors
do not care that the math came from Claude. They care that there is a
spreadsheet, printable, traceable, with a header containing the lab's
KOLAS / A2LA / UKAS accreditation number. The audit happens in Excel.
Always.

So: free for the curious, paid for the operational.

The numbers honestly

I am running this as a 14-day public verification. The verdict is on
June 5:

  • 1 or more sales → niche hypothesis holds, build the $99 cert generator
  • 0 sales → pivot to direct outreach (cold DMs to lab QA managers)

I will write the result either way.

What I have learned in week 1

Lesson 1 — The toolkit is for ~10,000 people, not 1 billion.
Every accredited calibration lab globally has to update its workflows
this year (ISO 10012:2026 was published in February). That is ~10K labs.
At $19 average sale, that is a $190K ceiling if I capture even 10% — and
I do not need to. I need 10-50 sales to validate.

Lesson 2 — The Excel toolkit and the MCP server reach DIFFERENT
buyers. The MCP attracts AI-curious engineers. The Excel attracts
ISO 17025 QA managers, mostly 40+ years old, mostly off Twitter.
Distribution channels do not overlap. You have to build both.

Lesson 3 — I delayed publish for 3 days. The product had been built
and uploaded to Gumroad as "Unpublished" since Tuesday. The 5-minute
"toggle to Published" step took 72 hours of avoidance. I think the
honest reason is: writing "the toolkit is for 10K people, not 1B" and
clicking publish are different actions. The first commits to the audience.
The second commits to the experiment that might prove me wrong.

I noticed the same pattern five years ago in audit reports — engineers
sitting on unambiguous data for two weeks. They were not lazy. They were
avoiding the moment after filing when the data either backed them up or did
not.

That is exactly what publish was.

The architecture, if you want it

  • MCP server: Python, FastMCP, hosted on MCPize. 10 typed tools. Reference: https://github.com/kyb8801/measurement-uncertainty-mcp
  • Excel toolkit: ten budget templates + a Welch-Satterthwaite + Student-t coverage factor calculator. Both English and Korean editions shipped same day.
  • Distribution: dev.to (this post), Medium (long-form), Beehiiv newsletter (weekly), Gumroad (transaction layer).

What I am watching for the next 14 days

  • Sales count (binary verdict at 1)
  • MCP server stars vs Gumroad sales — which channel grows the other?
  • Cold DM reply rate from KOLAS / A2LA / UKAS lab QA managers (parallel campaign, in case Gumroad discovery is zero)
  • Whether the $99 cert generator alpha (already built) gets even one pre-order interest from this batch of buyers

I will write the verdict on June 5.

If you are an MCP author shipping companion products — or an ISO 17025
auditor curious about MCP — leave a comment. I am collecting use cases
for the next iteration.

Top comments (1)

Some comments may only be visible to logged-in visitors. Sign in to view all comments.