skoshx

So for the past month or so, I've been working on a browser-based AI agent called Flytrap for automating the most dreadful part of software dev; fixing bugs.

It's completely free to use, so you can try it out, no signup needed here: https://useflytrap.com/ai

I wanted to ship it as fast as possible, to get real feedback how it performs for real world dev tasks, so all feedback is super appreciated! :) You can submit a task in the product, if the agent doesn't perform a task well, and I'll try to use those submissions to improve Flytrap

It contains 2 agents LLM Agent: This one is meant to be the bug-fixing AI agent, but I need to fine-tune it more to become better. It can perform tasks like read and write FS, execute shell commands etc. Simple Agent: This one is quite practical, it can write code, and then verify that it works as expected, so you can use it for code-snippets, that are verified to work (this eliminates hallucinations)

Some features - Browser based, so all your files are only accessible by you, saved on your localStorage - Clone from GitHub supported - Free to use

Next up I'm working on: - Making the agents more robust (fine-tuning) - Linear & GitHub integrations: one-click GitHub issue or Linear issue, agent clones repo and starts working on the task. - Browsing ability

Would love to hear brutally honest feedback

Join the Flytrap Discord to give feedback :) Thank you so much! - https://discord.gg/tQaADUfdeP

A blog post that shows how we implemented Rust inspired error-handling at Flytrap to solve many problems of good error handling, and what open-source JS libraries we use to achieve this.

Pentagon is a Prisma-like ORM built on top of Deno KV. Allows you to write your database schemas and relations using Zod schemas, and run queries using familiar syntax from Prisma.

I really like the simplicity, and lack of any build step or extra tooling.

I'm considering adding support for other KV stores such as Redis, if someone would find that useful :)

A blog post where we discuss how we do Rust inspired, human-friendly error-handling at Flytrap to solve many problems of good error handling, and what open-source JS libraries we use to achieve this.

Hey, thanks for sharing your perspective! While I do agree with your point to some extent, I somewhat disagree because product analytcs are usually a non-invasive way to actually improve the product for the customer, whereas most people use adblockers to block invasive things such as popups, ads etc.

With new browsers like Arc blocking ads by default and users flocking to them, it's getting tougher to get the full picture about how your product is used from product analytics.

This was something we noticed while building Flytrap, that there were many datapoints missing, which made it impossible for us to understand things about the product and what was going wrong etc.

This is a short blog post on how we use Next.js redirects to go around all ad-blockers!

Hope it helps product people on here! :)

- @skoshx

Good log/error messages are SO UNDERRATED. They guide your users to the solution, minimize support calls, improve customer and user satisfaction & build trust.

That's why I built `human-logs`, inspired by Vercel's Error framework. It allows you to take events, explanations and solutions and connect them like lego-pieces to create user-friendly logs.

Check it out on GitHub: https://github.com/skoshx/human-logs Or a demo: https://human-logs.vercel.app/

When building a quite a complex product like this, there were many interesting technical problems to face - for instance; creating the code-transform to transform code in a way that works in all libraries was quite tricky because we had to accomodate to all the intricacies of for instance `this` keyword in JavaScript. Another interesting problem was string serializing everything, and deciding what not to serialize (like for instance class instances)