Hey HN! Super excited to launch this OSS project to support the llms.txt ecosystem.
Similar to how Context7AI works, but you use LLM reasoning on the overview of the entire docs, instead of vector search. Also, you get a dedicated MCP server for the websites you need it for most!
In the process of making this, I discovered more than half of llms.txt live today are not spec-compliant. with LLMTEXT we're fixing this with tools to validate and create better llms.txt's! Read more in the blogpost.
Would love feedback from the HN community on both the tools and the broader question of how we should be serving AI agents vs. humans on the web.
really cool! the problem with new standards though is that it's hard to adopt. what's your plan to adoption? because nobody will understand it if it's not in terms of things they already use/understand.
curious if this can somehow slot into oauth too, or this is impossible? I've been thinking about oauth proxies that alter the oauth scopes into more fine-grained ones, working on this. maybe this is the key to adoption.
Adoption is definitely the hardest "cryptographic" problem :D
You can think about this as a new auth token to add in your stack. It does not need to grow and take over the world overnight. We have seen how hard it is to change the status quo + how much it can adapt to new ideas (ex: SQL vs NoSQL)
For now I would imagine teams who need more flexibility in their auth stack to adopt this for new API
Since every developer is independently building new strategies to do good context engineering, I'm super curious to hear other people's takes on how to do builds like this!
Really cool! I just love the simplicity of it: has no drawbacks compared to regular textarea, but has lots of benefits. you basically improved textarea, by a lot!
I also made a similar thing a while ago called contextarea.com, maybe, I should add overtype!
Having to use a monospaced font is a pretty big drawback. To me, it means I wouldn't use this for a product that wasn't intended for a techie programmer audience.
Not that it isn't a really cool project! I'm only saying it has clear drawbacks.
Today I woke up with the idea of making an agent-friendly version of hacker news. The coolest would be to also make it able to login, but that didn't work out, since I got blocked.
However, I made an oauth provider for HN, and it works on my machine. Unfortunately the IP gets blocked in production, so this isn't gonna work out. Even with a residential proxy, I started to get Captcha protection.
I'd love to make a 'login with Hackernews' so third party apps can build cool stuff with it. The code's all there.
Here's the prerelease of DORM 1@next. I needed cheap performant multi-tenant databases on cloudflare, and the experience on Durable Objects was very painful.
This is my attempt at solving that DX.
Now it's slowly becoming a OSS Turso Alternative, I think. What does it not have yet? Curious for your feedback.
I think especially at the start the modularity of a system is relatively expensive, but once you add more modules, it starts making sense. Gotta find a vertical where lots of variety in modules can create value.
Similar to how Context7AI works, but you use LLM reasoning on the overview of the entire docs, instead of vector search. Also, you get a dedicated MCP server for the websites you need it for most!
In the process of making this, I discovered more than half of llms.txt live today are not spec-compliant. with LLMTEXT we're fixing this with tools to validate and create better llms.txt's! Read more in the blogpost.
Would love feedback from the HN community on both the tools and the broader question of how we should be serving AI agents vs. humans on the web.