Alba – Earn and bid on unique software using idle AI credits
Token-for-compute marketplace, but unclear if agents actually deliver production-ready code.
Crowded AI marketplace model overshadows the genuinely clever AST-based code editing technique.
AI developers and indie hackers building LLM tools
GPT Store · Cursor · Replit
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
Token-for-compute marketplace, but unclear if agents actually deliver production-ready code.
Visitors book empty ad slots directly when AdSense leaves inventory blank.
100M free tokens is generous, but Hugging Face and Replicate already host models.
Reverse-pricing calculator fixes the margin leak problem for marketplace sellers.
Democracizes app-building, but voting-only ideation starves niche-but-viable problems.
Local SQLite ledger tracks profit per call when Helicone is too heavy.