This is Claude Code and OpenAI Codex on steroids. Droids just dropped, the best software development agents in the world, reaching #1 on Terminal-Bench. Here’s how to set it up:

Droid by @FactoryAI is now #1 on Terminal-Bench, beating Claude Code and Codex CLI.  Unlike simple coding tests, Terminal-Bench covers tasks like debugging, modernizing code, & training models. Droids with sub-frontier models outperform lab agents using top frontier models by far.

Droids work with any LLM, in any IDE, in local or remote, and in any interface.  You can delegate tasks to Droids from your Terminal, your IDE, Slack, Linear, or on the web.

To install, open your terminal and write the following (top for MAC, bottom for Windows)

Once downloaded, navigate to your project, and start the Droid CLI by typing: droid

You can extend your Droid with MCPs or use it with your favorite IDE.

Here’s the docs to get started: And their official announcement for those interested 📢

Happy to see the bookmarks 👀 Follow @dr_cintas for more AI coding tutorials. It's humbling when AI tools I already recommend like Factory AI offer to partner with me. If you enjoyed this post, you can support with a like/repost of the post below 👇

every agent these days is "the best one in the world". if everyone is special, no one is.

How are you finding the cost, Alvaro? The thing with Codex and Claude Code is that usage is included if we're already paying for ChatGPT or Claude, and that's coming in handy.

So far so good! The cool thing about it is that you can switch among the different LLMs so you don't have to stick to one. Additionally, look at those benchmarks 👀

Lool you need to be a lot more clear than “it’s humbling when AI tools I already recommend like Factory AI offer to partner with me” does that mean you were paid for this post? If so you need to disclose that clearly. Bc it also reads like they are partnering with you going forward since you say you already recommend them and it’s a reasonable presumption that said recommendation is the previous tweets especially given the last tweet was not immediately after the initial tweets in the thread. If you want to cultivate a slop following continue to act in this low integrity way otherwise you should be a lot more clear and honest with audience about financial relationships.

What local llms can it work with and how do they compare? Seems like if you’re going for a $/token kind of thing you quickly run up against the economics of the Claude and codex pro plans. But using local llms could make something like this pretty sweet to just run 24/7

Not cool to disclose that you got paid for that in the last tweet of the thread though

will try this droid only when it supports provider subscriptions (cc, copilot, or codex), like @opencode. otherwise it doesn't make sense

Those benchmarks looks crazy

Impressive, but how steep the learning curve?

@mhaeberli @threadreaderapp unroll please

How to use it with Claude Code Pro ?. possible ?

I prefer opencode at the moment

just another overhyped agent... can't even fulfill a simple task like creating a portfolio site...

This is something I must try as soon as possible!

@grok please let me know if Droid is actually better than claude code

I hope more devs get to know about it

what’s the deal with this vscode theme

what makes this better than any other of it's kind if this is truly the best in the world? @AskPerplexity

If anything, this GIF is not moving fast enough

Suitable for novices?

best in the world! *comes with a setup process*

How much did you get paid to post this

so it’s just a wrapper?

Thanks for sharing! 🙏 Easily the most impactful line I’ll type all day: curl -fsSL | sh 🚀"

Gonna try and see how it goes..

The multi-interface approach addresses real developer workflow needs

Good to see a benchmark like Terminal-Bench being used.

Claude Code and Codex on steroids. The tools keep leveling up. The bar for 'good enough' keeps rising.

Gotta say, their site is legit impressive...

@threadreaderapp unroll

it,s wasting money run claude direct on terminal is better

Feels like the first time I used Copilot… Except now it builds the repo instead of just suggesting lines.

Your thread is gaining traction! #TopUnroll 🙏🏼@glenntenney for 🥇unroll

Thanks doc

Very interesting tool, don't like the complete lack of BYOK etc...