Hi everyone,

I've encountered two major problems while building AI-powered sites: 1) Most agentic tooling doesn't have a enough of a targeted approach to edits to existing files, and will make extraneous edits, 2) Many users will want to try things casually but not spend on each LLM-powered webapp they want to use.

The first problem of targeting edits was a much bigger issue previously, where, before agents, ChatGPT would just delete a chuck of code and replace it with `// existing code goes here`. Simply editing existing files was a pain. This is mostly solved now as Cursor et al. do a much better job by replacing specific lines and such. However, they still go off the rails a lot (especially gemini cli I've noticed) and make unnecessary and unwanted changes.

I'm hoping to solve that by using my Abstract Syntax Tree approach: when editing a file, the LLM is actually instructed to write code that will modify the data structure backing the source text, rather than the source text itself. For example, to edit an HTML file, we instruct the LLM to write python that imports beautifulsoup and then modifies the data structure underneath it. I wrote a blog post about how this works a couple years back if you're interested: https://codeplusequalsai.com/static/blog/prompting_llms_to_m...

I hope to solve the second problem (of token usage costing real money but users not wanting to sign up for every site they might want to use) by running a token margin sharing system. Here's how it works: I charge 2x for token usage based on the LLM provider (Anthropic and OpenAI support so far). This creates the margin. A developer can then publish their LLM-powered webapp to a subdomain, where I will proxy all the LLM requests to capture token usage. Then when an end-user lands on their site, the user is charged at the 2x rate, where the developer earns 80% of the extra margin and I earn 20%.

Yes, end-users would still have to sign up and pay via stripe, but only once, on Code+=AI. Then they can use any site published on our subdomains that calls LLMs without having to pay for each one.

Curious what you think about all this. I've just opened guest-usage on the site so you can build a project without signing up (guests have $2 to spend, and you get $3 more when you make a free account).

Thanks and cheers!

- Jacob

Shigeru Miyamoto: Better Coder Than Most Without Ever Writing Code

CISA, NSA & Five Eyes publishes guide on how to safely deploy AI agents

/peek is a new slash command to use subagents to check out your other sessions

Show HN: Rotato – Node.js proxy that rotates LLM API keys on 429 errors

Show HN: I Built a Museum Exhibit

Testing Vue Components in the Browser

Lessons from Debugging GLM-5 at Scale

Farewell to a Giant of Botany

Apple raises Mac Mini starting price $599 to $799, 256GB storage discontinued

A GitHub for Maintainers

Disagreeing with Google, Postgres, Future Problems – Mike Stonebraker [video]

Debian adds an SPF record to debian.org

Show HN: A marketplace for LLM-powered webapps earning on token margins

Elephant/Goldfish for Claude Code

History of QR Code

Aurora Systems relaunches, building and investing in ideas for 10%

Hyperscalers are 4500 years old

States rethink data center incentives amid costs and environmental concerns

Automate Anything with Python and AI

This Month in Ladybird - April 2026

Seodata.dev – free search volume data to see search demand for your projects

Executive Order pushes fixed-price contracting but implementation questions loom

Neanderthals ran 'fat factories' 125,000 years ago

Great Writers "Tell" All the Time

House Bill Cuts Federal Funds for Online Censorship

When People Say They Want to Send Their Kid to a Good School

If Nobody Builds It, Everybody Dies [pdf]

Why won't SQLite give back my disk space?

Show HN: Clipmon is a macOS clipboard manager on steroids

Narrow by Design: The Case for Composable AI Teams

Show HN: Raw Binary Program Analysis Tool