Two weeks ago I made leantext public: a free, open-source tool that turns web pages, PDFs, screenshots and Office files into lean Markdown before your coding agent reads them.
In my last post I shared the launch-week numbers. This week something better happened: it kept growing without me posting anything.
The numbers
- 73 git clones from 49 unique cloners
- The README passed 100 views, from 23 people
-
3 people opened
install.shto read it before running it, which is exactly what you should do with any installer
Where people are coming from
This is the part I find most interesting. In launch week, almost everyone came from my own posts. This week the traffic came from places I never posted:
-
GitHub itself: 6 people found leantext through GitHub search, topic pages or someone's feed. Adding the
ai-agents,claude-codeandllmtopics to the repo turned out to matter. - Google: 3 people searched their way to it.
- Reddit: someone shared it, and it's still sending visitors.
- X and LinkedIn: a couple of people each, days after the posts went out.
Organic discovery from Google, GitHub and Reddit is the signal I was hoping for. It means people are looking for this problem and finding leantext when they do.
A quick recap, if you're new here
Your agent pays for every token it reads, including navigation, sidebars and hidden markup it never needed.
| Input | Before | With leantext | Saved |
|---|---|---|---|
| PyPI: NumPy | 92,348 | 755 | 99% |
| PyPI: uv | 43,177 | 2,501 | 94% |
| Stripe API docs | 3,958 | 1,194 | 70% |
| Attention Is All You Need (PDF) | 10,168 | 9,838 | 3% |
That last row is there on purpose. leantext can only remove noise, so a research paper that's all content barely shrinks, and that's the correct result.
In end-to-end Claude Code runs, 21 out of 21 tasks stayed correct with leantext on. A Claude Code hook converts the agent's own Read and curl calls automatically, and if a conversion looks wrong or saves less than 15%, the agent just gets the original. You never get a worse result than without it.
A new banner, made with living-canvas
The repo has a new animated banner that tells the whole story in three panels: 92,348 tokens of PyPI clutter, leantext checking the conversion, and 755 clean tokens coming out.
I made it with living-canvas, my Claude skill that turns any repo into an animated README banner. More on that one soon.
📺 Now on DEV·TV Spotlight
leantext is now featured on the Spotlight channel of DEV·TV, my free browser TV for developer content. Flip through GitHub, Hacker News, DEV, new models, agent skills and more, like channel-surfing in the 90s. No login, no backend.
Want your own project on air? Spotlight takes submissions right from the TV.
Try it
git clone https://github.com/shouvik12/leantext && cd leantext && ./install.sh
About 20 seconds, 39 MB on a Mac. Say yes to the Claude Code hook and keep working as usual.
The most useful thing you can send me is a bad conversion: the page, the output, and the task that failed. Every one becomes a test in the repo.
👉 github.com/shouvik12/leantext
Free, open source, Apache-2.0. Thanks to all 49 of you who cloned it. 🙏
Top comments (0)