Can anyone from vlt explain the inconsistencies in their performance over time graph? When I compare it w/ the npm baseline, it looks like it fluctuates between being ~4s faster typically but then there are periods where the performance is worse.

I appreciate the transparency that the company has showing their performance, the early August degradation has me wondering what’s going on.

I believe most, if not all of those, were due to some variation of adding some new feature/capability to the system and then making things faster again.

Anyone use this and have also used cmux?

Loved the article, thanks for sharing. I’m curious if you’d share your setup. I haven’t made any macOS apps before, primarily because I never wanted to really learn XCode and obj-c. I like swift but still prefer simpler editors like Zed/VSC vs. What XCode offers.. so when you’re building these are you doing it in XCode or in another tool like Claude Code/codex/gemini CLI?

cheers

Seconded, I would be interested in knowing people's workflows + experiences developing MacOS and iOS apps with claude, etc.

From the repo here, it looks like its just using swift command line tools, which might just work well enough with cursor/vscode/etc. for small projects. You won't have Xcode's other features but maybe thats fine for an agentic-first development workflow.

Not OP, but I use Xcode with Claude Pro and it is going fairly well. I also am creating my own personal-use apps instead of paying for monthly subscriptions. I know a bit of Swift, and had been trying to learn it while also using LLMs. At this point, I've decided to also not make these real projects and just vibe-code exactly what I want.

I've been using Cursor with the Swift Extension. Works really nicely, but sometimes I switch to Xcode to do some tasks such as testing that it works to build.

Can you use beads with other task management systems like linear or shortcut?

Also does beads work with obra’s superpowers plugin?

No idea, I know its open source and the LLM manages it directly. I tell Claude what to add etc and it does it.

I didn’t realize you can have subagents have different sets of MCP servers+tools enabled & disabled. Do you have any docs for this? A slash command to trigger different types of workflows would work well.

Between this and models like exaone 4 32b reasoning is it possible to have a TUI setup similar to Claude code?

Amp Code is also very good, they released about 2 weeks ago their Oracle feature which leverages o3 to do reviews (https://ampcode.com/news/oracle). Amp leans the closer to Claude Code more than other solutions I’ve seen so far, the team there is really leaning into the agentic approach.

I watch the changes on Kilo Code as well (https://github.com/Kilo-Org/kilocode). Their goal is to merge the best from Cline & Roo Code then sprinkle their own improvements on top.

There’s a lot of talk about Claude Code in here, and I agree it’s a great agentic coding tool. One of the benefits of Cursor & Windsurf is/was the ease for smaller companies to setup Team accounts and have control over spend.

Claude Code I think misses this. You can get an enterprise account if you commit to over, what.. 70 seats annually?

If you’re an individual you can get Max 5x/20x ..

But for smaller companies, I don’t think they are addressing that space. Am I wrong? Are there any Agentic tools like Claude Code that can provide a fixed cost per user?

Enterprise is where you make the money. I’m sure they will have something soon.

For now, i can just piggybacking on bedrock or vertex which is peanuts comparing to production cost.

> piggybacking on bedrock or vertex Can you elaborate? I’m not familiar with this approach.

A small company can just pay the $250 a month for X number of employees to each have CC max plan. Not that complicated

I thought that was against the terms of service from Anthropic and they have the rights to disable/ban accounts doing that.

agree. With the fall update to Claude 3.5 Sonnet it’s held its own. Add in MCP and it’s pretty good. Looking forward to seeing what they’ve achieved.

Like psychoslave suggested, try out mise (https://github.com/jdx/mise). I used asdf for years, did the switch to mise and have never looked back for package management. It supports a huge number of languages and is performant.