AI & Agents
Holt Cost ₹6,000 and 100 Hours. It Became One of 30 VibeCode30 Winners.
No funding, no team, built with Claude Code. What it took to ship Holt, an open-source memory layer for AI agents, and how it ended up in the official VibeCode30 2026 report.
VibeCode30built with Claude Codesolo founder AI project
The official VibeCode30 report has a page for Holt, and it carries two numbers I find hard to read without laughing a little: 100 hours and ₹6,000. That is what the report lists for what it took to build it. No funding, no team, and Claude Code doing a lot of the typing.
On 14 October 2026 I stood on a stage at YourStory TechSparks in Bengaluru as one of the 30 VibeCode30 winners. This is the plain account of what that project was, what the numbers do and do not tell you, and what I think a solo builder can realistically ship now.
What is VibeCode30?
VibeCode30 2026 was a showcase of 30 products built with AI coding tools, run at YourStory TechSparks in Bengaluru. The official report calls all 30 featured builders winners, so the accurate wording is that Holt is one of the 30 VibeCode30 winners. I am careful with that phrasing because “won VibeCode30” would imply a single first place, and that is not what happened. You can read the official VibeCode30 Report 2026 yourself.
The report’s page for Holt gives this line: “Every AI session starts from zero. Holt is the part that remembers.” It also lists 5,207 downloads by late September, 100 hours of work, ₹6,000 in cost, and Claude Code as the tool it was built with.
What is Holt?
Holt is a local memory layer that sits under whichever AI agent harness you already use, such as Claude Code, Codex, Gemini CLI or Cursor. It is not an agent and it is not infrastructure. Hooks save and recall memory around each session, MCP exposes recall and remember as tools, and the data is plain local files with local embeddings.
A few properties I care about: memory is kept per folder, there is a trust gate on what gets written, forgetting is real and not just a hidden flag, and the licence is MIT.
On the numbers side, Holt 0.19 scores 91.2% on LongMemEval, 88.2% on LoCoMo and 79.6% on BEAM at 1M tokens, rerunnable in a public benchmarks repo. I wrote up how I decide which Holt numbers I publish separately, because I had a few wrong ones in my head before I checked.
How much did Holt really cost to build?
The report says ₹6,000 and 100 hours. I will say what I can stand behind and flag what I cannot.
What I can say: those are the figures printed in the report. What I will not do is invent a line-by-line breakdown of where the ₹6,000 went or which of the 100 hours went where, because I do not have one written down that I would trust. If I reconstructed it now, it would be a story about my own work, not a record.
So treat both figures as an order of magnitude. A working open-source product with thousands of downloads, built alone on a small budget. A few years ago I would have guessed that sentence needed a team.
Which part of building with Claude Code was easy, and which part was slow?
I need to be honest about the limits of what I know here. I do not have a time log that splits the 100 hours by task, so I cannot give you a tidy chart of “20 hours on X, 40 on Y.”
My honest read, as opinion and not measurement, is that the writing of code was the part AI compressed most. The parts that I suspect still ate the most clock time are the ones no tool does for you: deciding what the product should refuse to do, checking whether a number is real before saying it publicly, and making the thing safe enough that someone else would trust it with their notes.
The benchmark episode is a good small example. Producing a 97% figure was easy. Working out that it came from a 101-question tuning run took a deliberate stop and a look at the source, and a model was never going to volunteer that. That kind of slowness does not show up in a “built in 100 hours” headline.
What can one person ship in 2026?
Based on this one project, and I am wary of generalising from a sample of one:
- A usable, installed, open-source tool is within reach of a single person with a small budget. 5,207 downloads by late September is proof that strangers will try it.
- The cost that drops fastest is production. The cost that does not drop is judgment: what to build, what to leave out, what you are allowed to claim.
- Visibility still has to be earned separately. Building was the cheap half. Getting noticed among 30 builders at an event is a separate job from building.
What I would tell another solo builder
- Write down your hours and spend as you go. I wish I had a log, and the report numbers would be sturdier for it.
- Keep a list of which of your claims you have verified and which you only believe.
- Use the exact wording of any recognition. “One of 30 winners” is true, and true is better than bigger.
- Ship the open-source version before you perfect it. Downloads told me more than my own opinion did.
I am still not sure what the best use of the next hundred hours is. My bet is on measuring the things I have only claimed so far.