Winding Down Tandemu

There’s no graceful way to open this one, so I’ll just say it.

Twelve days ago I published a post about shutting down Substaff. In it, I named Tandemu as one of the two projects I was redirecting my attention to — “the one I keep thinking about when I’m not at my desk.” Two months before that, in the Kreye post, I put Tandemu at the top of the same list.

Today I’m shutting down Tandemu.

If you’ve read all three posts in order, that’s a pattern that deserves an explanation rather than another round of the same paragraph about infrastructure costs. So let me give you the real one.

What Tandemu was

Tandemu was the management layer for AI-assisted software development. Developers used Claude Code the way they already did; Tandemu sat underneath and quietly collected session telemetry, then turned it into the things engineering leads actually want to know — how much of the codebase is AI-written versus hand-written, where people get stuck, how work actually flows — without standups, timesheets, or anyone filling in a form.

It also carried a persistent memory layer, so the AI teammate remembered how each developer worked from one session to the next instead of starting cold every morning.

The pitch was “stealth observability”: session-level insight without surveillance. Metadata only, never keystrokes or code content. I still think that framing is right.

What actually changed

In August I stepped into a bigger role at dubizzle — Associate Director of Engineering. That is the honest, unglamorous reason the calendar stopped working.

I had been running three products alongside a demanding day job. That was already tight. With a wider remit at work, it stopped being tight and started being fictional. Something had to go, and the question became: if I can properly support two, which two?

Tandemu lost that comparison, and not because it was the weakest idea. It lost because of what kind of product it is.

Tandemu is B2B, and B2B needs a motion I can’t give it. AnswerRank is self-serve — an indie founder finds it, signs up, and gets value alone in a browser tab. FoodBang is consumer — it lives or dies on whether cooks and diners show up. Tandemu needed an engineering lead to adopt it, then get a team to install a plugin, then keep them installed long enough for the metrics to mean anything. That’s a sales-and-onboarding motion, and a sales motion is the one thing you cannot do in the margins of a full-time job. The product was fine. The distribution required a person with time I no longer have.

And the usual arithmetic applied on top. An always-on telemetry platform bills you whether it’s ingesting from fifty developers or zero — database, ingest pipeline, hosting. For low activity, that’s a monthly donation to a cloud provider.

What I got wrong

Two things, and the first one is more interesting than the cost story.

I built a measurement product for a workflow that was still moving. Tandemu measured how developers work with AI assistants. But how developers work with AI assistants changed substantially over the life of the project — new tools, new modes, agents doing multi-step work rather than single completions. Every few months, the thing I was measuring reshaped itself. A metrics product needs its underlying reality to hold still long enough for a baseline to be meaningful. “Your AI ratio is 47%” only lands if 47% means the same thing this quarter as it did last quarter. I was drawing dashboards on top of a moving floor.

I assumed leads wanted visibility more than they wanted output. Tandemu was built on the premise that engineering leads would pay to understand AI-assisted work. What most of them actually wanted was more of it — ship faster, not measure better. Visibility is a second-order want. It becomes urgent once AI-assisted work is large enough and contested enough that people are arguing about it, and for most teams it isn’t there yet. Same mistake I made with Substaff, in a different costume: I built the instrumentation layer before the thing being instrumented was big enough to need instruments.

Noticing that I’ve now made the same category of error twice is the most useful thing I’m taking out of this.

The code stays open

Here’s the part that’s genuinely different from the other two shutdowns.

Tandemu was source-available from the start, and it stays that way. The hosted platform at app.tandemu.dev is gone, but the self-hostable core is still there under the Elastic License 2.0 — free to use and self-host, though not to resell as a competing hosted service — and self-hosting still works end to end with Docker Compose:

• Repo: github.com/sebastiangrebe/tandemu
• Docs: the self-hosting guide on tandemu.dev stays online

If you’re building anything in the developer-telemetry space, there’s real material in there: a CLI plugin system for Claude Code and OpenCode, a session telemetry pipeline that stays on the right side of the surveillance line, DORA metrics computed from real signals rather than ticket hygiene, and an AI-attribution approach that doesn’t require developers to self-report. Take what’s useful.

What I’m actually keeping

Two, and this time I’d rather be held to it:

• FoodBang — a live video feed for street cooks, home kitchens, and tiny spots that open only when they’re actually cooking. Post a 15-second clip, open shop for the day, take orders in DMs. No fees, no middleman.
• AnswerRank — visibility in the era of answer engines. Track how ChatGPT, Claude, Perplexity, and Gemini talk about your brand, then work a concrete checklist so they recommend you instead of a competitor.

One consumer, one self-serve. Neither needs me to run a sales process to work.

A thank you

If you installed the plugin, connected a repo, sent me a screenshot of your AI ratio, or told me a metric was wrong — thank you. The people who used Tandemu were engineers with strong opinions about being measured, and the feedback was correspondingly sharp and useful. That’s the good kind of user to have.

I’ve now written three of these. What I’d tell anyone reading them in sequence: the shutting down isn’t the failure. Running four half-supported products while telling yourself you’ll get to them is the failure. The post is just where I stop pretending.

Two projects. Enough hours to actually do them. On to that.

If you want to follow where I’m headed, the projects page is the live list, and you can always reach out.