miketromba

There’s a horizontal body overflow bug on this page on mobile. The irony

Real human-in-the-loop is not supported, cronloop is designed for use cases that benefit from full agent autonomy.

You can, however, continually monitor and steer the agent by refining its instruction set / memory to improve it over time so that it makes better decisions itself.

I have claude connected to the Cronloop MCP and occasionally ask it to analyze the previous N runs of my agent to see if its instructions should be adjusted.

/loop is sufficient in many simple cases

Cronloop is more of a control plane that makes it nicer & more scalable once you have many agents and need to monitor / refine them over time.

Create & manage unlimited agents in one place, they all run in the cloud - no machine to manage and keep online yourself, easily monitor and analyze runs (agent runs stream in real-time), tune the agent with run duration constraints, give it the ability to self-improve with the built-in durable memory system without having to invent your own, connecting tools is very simple (200+ supported in the connector library).

It also runs codex, not just claude code.

I currently have ~25 cronloop agents running on various models between codex/claude.

Running them all on /loop would be painful and hard to maintain without cronloop or a similar control plane.

Hey HN, I often want to automate recurring tasks with my Claude Code/Codex subscriptions, so I built Cronloop to make this simple.

Just create an agent, set a schedule, connect your existing claude code/codex subscription, give your agent instructions + tools, and boom - you have an autonomous agent running in a loop, working while you sleep.

Every agent has a simple, durable markdown-based memory system. Your agents can self-improve over time - you can include instructions to record learnings with each run for other agents to benefit from.

Some ways I'm using it:

- I built a self-driving events website (aievents.now) and have a cronloop agent for each city that autonomously curates events every morning. - I have a few websites autonomously SEO-optimizing every day based on search console data + keyword research.

Free to try - you can create up to three agents for free to get started.

Hey HN, I often want to automate recurring tasks with my Claude Code/Codex subscriptions, so I built Cronloop to make this simple.

Just click "New agent", set a schedule, connect your existing claude code/codex subscription, give your agent instructions, connect the tools it needs, and boom - you have an autonomous agent running in a loop, working while you sleep.

Some ways I'm using it:

- I built a self-driving events website (aievents.now) and have a cronloop agent for each city that autonomously curates events every morning. - I have a few websites autonomously SEO-optimizing every day based on search console data + keyword research.

Some features:

- Every agent runs in an ephemeral, isolated sandbox - You can monitor runs in real-time (the agent logs stream into the app) - Every agent has a simple, durable markdown-based memory system. Your agents can self-improve over time - you can include instructions to record learnings with each run for other agents to benefit from. - You can leverage your existing claude code/codex subscriptions or bring your own API key. - 200+ connectors are ready-to-go - connect your agents to any tool that has an MCP, CLI, SDK, API, etc. - You can configure the agent's environment with a setup shell script if needed for greater flexibility

Free to try - you can create up to three agents for free to get started.

You bring your own inference, so cronloop is very generous with usage, the pro plan offers unlimited agents with up to a 5 min interval on runs for $25/mo (or $20/mo annual).

I'm a huge Cronloop user and get a lot of value from it, so I'm excited to share it. Let me know your thoughts!

*Hand-written comment, no AI

This is very good - I have found that well-designed progressive disclosure systems like this tend to be far more effective and performant than cookie-cutter semantic / encoding-based RAG systems. Agents seem to love tables of contents.

I get this. Just shipped the ability to create and edit themes locally, no auth required. The local theme gets persisted to localstorage and you can optionally save/share it later. It also works seamlessly with the fork / import features, so those can be used without auth as well.

Just shipped a fix, middle-click should work now. Thanks for the feedback

Global border radius is editable, that setting is at the bottom of the sidebar. The challenge with global shadcn theming is that you're limited to adjusting the css variables they provide. I believe there is a global spacing variable, but it is not so specific that you can target e.g. just label spacing. That would be something you could modify directly within your shadcn input components via adjusting the tailwind class(es).

Tweakcn is a great tool too. Main difference is I'm hoping ShadcnThemer will be more of a community-driven hub for sharing, starring, and forking themes - similar to how color palette websites have 1000's of user-made palettes. (I took this approach when building the Theme Studio for VS Code and it worked really well, 1000's of themes were designed and shared.)

Tweakcn also charges $ users to be able to share and save themes which I think is silly for a tool like this, should be 100% free and open source.

I also prefer the simple UX of ShadcnThemer better but I'm biased of course.

Excellent work. A modern alternative to readability was much needed. This is especially useful for building clean web context for LLMs. Thanks for open-sourcing this!

Oof this is a very limited take.

The fundamental issue is that agents / LLMs at present are not engineers or system architects. You cannot sit back and play product manager yet. If you go in with this expectation, you will certainly have a poor experience.

LLM-powered coding agents are not comparable to any existing human role or tool and therefore cannot be used interchangeably. They are a fundamentally new and novel type of tool that enable incredible productivity IF and only if you use them effectively. The tool is not the problem, the user is.

I’m also building a non-trivial SaaS platform and the cursor agent has written 90%+ of the code in the codebase. It has architected ~0% of the codebase. It has modeled ~0% of the data layer in the codebase. It has enabled me to move at least 10x the speed of implementation and debugging (yes, it is incredible at debugging if you use the right techniques).

I use the word implementation because that is almost exclusively what I use it for. I do the planning, design, architecture, engineering. Then, I communicate those specifications to it in digestible chunks, and it implements them incredibly fast. It makes mistakes sometimes, comes up with poor naming conventions, fails to reuse existing modules in the codebase, etc. But that’s fine - I simply give it feedback, point it in the right direction, and within seconds it’s back on course.

I own the codebase, it’s just my super fast implementation minion. If it gets something wrong - that’s my fault. I failed to communicate my expectations clearly. I’m heavily using my brain while working with the LLM - I’m just reserving those mental cycles for higher-level decision making than the actual tokens to type into the editor.

The power is in your ability to steer it and communicate in clear terms what you want, and maintain the right level of abstraction in your instructions. Too specific and you lose some of the benefit beyond just writing the code by hand, too vague / high-level and it will over-engineer something totally different than what you intended.

You’re the engineer, it’s just a really smart semantic code generator. That might change in the future, but for now, if you use it in this way, the productivity benefit is very clear.

I’ve been using cursor with 3.7 sonnet max (now sonnet 4) and I already can’t imagine a world without coding agents - it is so deeply rooted into my workflow.

Thank you! Indeed, it's pretty early and desktop-first for now. Agreed- I intend to add lots of examples over time. Appreciate the feedback

Agreed, added a wall of disclaimers to clarify what it is and what it isn't.

Good call, added some more details.

Going to implement an "import from json" feature soon... that would let you do this

Third time I'm hearing this... I think it might have something to do with: a) newish domain, b) uncommon TLD ".one" or c) cloudflare DNS setup pointed to netlify with cname flattening. I'm a noob when it comes to DNS stuff, so I'm not sure if/why that would cause problems, but who knows. If anyone else has an idea, let me know!

--- I think I figured it out. Someone mentioned "Cisco Umbrella" blocking it. This article explains the categories of 'threats' it blocks: https://support.umbrella.com/hc/en-us/articles/115004563666-...

One of the categories is "newly seen domains" and another is "dynamic DNS". I suspect one (or both) of those is the culprit because everything else appears clean on total scans like this: https://www.virustotal.com/gui/domain/themes.vscode.one/dete...

I actually experimented with this in earlier iterations of the Theme Studio, but I remember running into a bunch of limitations. I think the HTML/CSS simulation is similar enough to the real thing for now (prob 90-95% accurate), but like you said, it would be cool to interact with the theme in a real editor. (clicking buttons, hover states, etc)

Good idea, noted!

Good idea. The screenshot generation system is a bit limited at the moment just because of the quantity of images that need to be generated on the server-side every time someone updates their theme. This does seem like an important one, so I will add this to my list, thanks!

Yeah that's a nasty bug. I built the picker from scratch so it has not been battletested. Will fix this moving forward.

Strange, do you know what's giving you that warning?

Thanks for sharing this, I will look into a light theme for a future iteration.

Yeah the Firebase + GitHub auth seems to be broken across-the-board. Going to dig into this.

Seems to be broken for now unfortunately; will have to figure this out.

That's fair. In the future I'd like to implement features like: automatically handle publishing to the marketplace for the user (right now it's a pain), upvotes system, etc. Accounts are a good solution for these types of features + it makes the theme data persistence way more robust so you can comfortably / reliably start a theme and then come back later to finish it. There are many suggestions for making signing-in optional though, so will explore this.

Yeah, this needs some curation / clean-up features for sure.

The Gem made me laugh. In game design there's some concept of '4 different player archetypes' and I forget the details but I remember one of them is like... that player that just likes to break stuff. Respect to whoever made that :)

Agreed, going to explore this, thanks!

I like Themer. I think Themer is awesome for creating less-detailed but still great 'umbrella' themes that you can bring with you to lots of different programs. This is for creating more detailed / robust VS Code-specific themes.

For sure, I would like to get a HN-type 'hot' algorithm with upvotes powering the list in the near future.