Myles hamilton

0

5

622

Myles hamilton@ProjectLancelot·3d

@_loganlee Let me know what you think when you try it. Happy to help if you hit any questions.

English

1

8

Logan Lee@_loganlee·4d

@ProjectLancelot My workflow could be covered it sounds like. I'll check it out.

English

0

10

Logan Lee@_loganlee·5d

“Anthropic killed OpenClaw.” This is all over my feed. My only questions: - can it run 24/7 workflows? - does it support a multi-agent setup?

You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.

English

7

1

19

751

Myles hamilton@ProjectLancelot·4d

@_loganlee Multi agent orchestration with desktop awareness. UAB lets agents read interface state locally without vision tokens, so your orchestration layer always knows what’s on screen. Desktop control is still sequential by nature, but it’s fast, free, and works with any model.

English

0

1

36

Logan Lee@_loganlee·4d

@ProjectLancelot Looks cool man, what's the main use case for your product??

English

0

19

Myles hamilton@ProjectLancelot·4d

Claude computer vision is slow and expensive to use. There are better ways to have desk top control. And they don't have to be tied to any specific agent frame work. I actually made a universal computer use local MCP server. No vision. Symantec control by using framework hooks. Essentialy reads the UI elements maps them to a 2d map that agents can read and act on. I call it UAB. Think of it like the usb port for desktop apps. Happy to answer any questions. It's free to use if you're interested or want computer use for your claws.

English

87

Johann Sathianathen@johann_sath·5d

openclaw is dead all because of claude computer use no claude runs purely on anthropic openclaw runs on your hardware, uses any model, improves itself & works when you're not at your computer it's like saying google docs killed linux

English

20

5

62

4.4K

Myles hamilton@ProjectLancelot·4d

Felix appreciate the transparency. Question: if Claude only sees one approved app at a time through the content filter, how does it handle workflows that span multiple apps? Like pulling data from one app and putting it in another? UAB reads the full desktop simultaneously without any content filter limitation. Genuine question about how Anthropic plans to solve cross-app workflows.

English

38

Felix Rieseberg@felixrieseberg·5d

I want to share a little more about how Claude using your computer works! When Claude wants to use your computer, it first needs to request access to specific applications (Slack, Figma, any other app).

English

72

42

825

77.1K

Myles hamilton@ProjectLancelot·4d

@EverydayAI_ @claudeai Jordan every one of those problems is the screenshot loop. UAB reads the interface directly. No screenshots. No session budget drain. No dedicated computer needed. Local MCP server. projectlancelot.dev/uab.html

English

0

1

57

Jordan Talks Everyday AI@EverydayAI_·5d

My 2 cents so far on this new @claudeai drop. 👇 - This is sloooooooow - If you can get over the slowness, it's pretty good - You have to remind Claude what it can't do. (Like, yes, Claude, you can copy and paste. lol) - If you wanna run this, it REALLY needs its own computer. Even running this on two monitors screen/mouse jacks your current progress - The real tell will be usage. I'm on the 20x max plan, and a VERY simple prompt/response took up 6% of my session budget. So, it wouldn't have even worked on the standard $20/mo plan.

You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.

English

6

3

27

3.4K

Myles hamilton@ProjectLancelot·4d

@Yuchenj_UW You diagnosed it exactly. UAB solves it today. Local MCP server. Reads accessibility tree directly and creates a 2d map of the whole UI. No screenshots. No remote vision model. Posted a tweet in under 3 seconds. projectlancelot.dev/uab.html

English

38

Yuchen Jin@Yuchenj_UW·5d

I used Claude Computer Use/Dispatch yesterday. My feeling: It’s too damn slow! Posting a tweet takes me ~5 seconds (once I have the content). Claude took 70 seconds. Why? It controls the screen via a loop: take a screenshot → send to a huge remote multimodal model (opus 4.6) → decide actions (click, type, scroll) → take another screenshot → repeat. We’re basically forcing a large general model to operate a human UI. Two things will happen in my opinion: 1. It is using a massive model (Opus 4.6) just to understand screens. That won’t last. Smaller, specialized models and eventually local models will handle most of this. 2. GUIs were built for humans. Almost all software will expose APIs/CLI for agents, so most actions won’t need to “use a computer” at all.

English

138

32

645

53.5K

Myles hamilton@ProjectLancelot·4d

@DavidOndrej1 Vision based computer use fails because it’s guessing from pixels. UAB reads the interface directly the same way you do. No screenshots. No coordinates. Works on Windows today. projectlancelot.dev/uab.html

English

52

David Ondrej@DavidOndrej1·5d

i tried most of these use-cases it failed on every single one.

You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.

English

157

35

997

115.4K

Myles hamilton@ProjectLancelot·4d

@Bhavani_00007 10/10 setup. Now imagine that watch telling your desktop to control any app without screenshots. UAB runs as a local MCP server. Your watch could orchestrate your entire desktop through Claude. projectlancelot.dev/uab.html

English

53

Bhavani.py@Bhavani_00007·5d

Rate my setup guyys

English

514

830

16K

830.9K

Myles hamilton@ProjectLancelot·4d

@jappleby Couldn't agree more! If they put in effort and the ideas are genuine. It's the subject matter that matters not the writing style.

English

22

Jack Appleby@jappleby·5d

So yes, ChatGPT writing is pretty easy to spot. But can I offer a different POV on AI writing? Most people aren’t very comfortable writing! And it’s scary to write, especially if you’re not confident in your writing skills. Not everyone was taught how to write. It’s actually magical that AI is helping people get their thoughts out more clearly, even if AI writing has some tells. I’d rather someone use AI to help share their thoughts with the world than not hear from that person. Maybe instead of spending time thinking people are ~cheating the system~ by using AI, we start celebrating how AI enables people to put together their ideas. Maybe we stop caring if something’s AI generated, give people the benefit of the doubt.

English

83

7

99

9.8K

Myles hamilton@ProjectLancelot·4d

@Jhaddix The supply chain problem is ungoverned agents. Lancelot is the architectural answer. Happy to answer any questions

English

126

JS0N Haddix@Jhaddix·5d

the litellm compromise is bad… But you’d 🤮 if you’ve seen some of the stuff in the AI supply chain I’ve seen 🫠

English

6

110

11.2K

Myles hamilton@ProjectLancelot·4d

Claude now talks directly to Grok inside the actual desktop app. or any desktop app. True computer use No APIs. No webhooks. No screenshots. It types, reads the response, replies , full multi-turn conversation through pure semantic control. This is what happens when you give agents a complete 2d spatial map of every in-code UI element so actions are instant and reliable. projectlancelot.dev/uab.html #UABBridge #ClaudeComputerUse” @grok @AnthropicAI

English

1

67

Myles hamilton@ProjectLancelot·5d

@mreflow Bet it can't do this!

English

75

Matt Wolfe@mreflow·5d

Initial impression after a little bit of use today... It's WAY too slow to be useful for anything other than asking it to do something remotely with dispatch. It has a mode where it will walk you through how to do something but it's painfully slow.

English

10

1

43

3.7K

Matt Wolfe@mreflow·5d

This looks pretty awesome.

You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.

English

14

5

100

14.3K

Myles hamilton@ProjectLancelot·5d

@felixrieseberg Mouse and keyboard get you there. Semantic control gets you there reliably. UAB reads the accessibility tree directly. No coordinates. No screenshots. No guessing. projectlancelot.dev/uab.html

English

1

111

Felix Rieseberg@felixrieseberg·6d

Today, we’re releasing a feature that allows Claude to control your computer: Mouse, keyboard, and screen, giving it the ability to use any app. I believe this is especially useful if used with Dispatch, which allows you to remotely control Claude on your computer while you’re away.

English

902

1.5K

18.8K

4.7M

Myles hamilton@ProjectLancelot·6d

@ZackKorman @Pier4r NemoClaw is a fence. Lancelot is an immune system.

English

31

Zack Korman@ZackKorman·6d

@Pier4r I mean that’s part of my criticism. I use AI to find this stuff. Just a good chat with opus and you find these problems. Not doing that is a choice

English

2

4

20

558

P@Pier4r·6d

This should be solved too. If LLMs are already 140 IQ, they can simply ask them, right?

Zack Korman@ZackKorman

NVIDIA fixed NemoClaw to "prevent the sandboxed AI agent from modifying gateway security settings (openclaw.json)" Except it didn't work. The AI can just make a copy of the settings and restart pointing at that new config. Same result. They're really struggling with the basics.

English

0

4

657

Myles hamilton@ProjectLancelot·6d

@ZackKorman @Setheroo88 The problem is no one is building security as the architecture its all being added in later instead of building from that perspective first. most solve one problem and ignore another.

English

2

14

Zack Korman@ZackKorman·6d

@Setheroo88 At this point I feel people just say they’re secure to get me to pentest their shit for free

English

11

1

79

3K

Seth Thompson@Setheroo88·6d

Zack is basically the final test for anything that claims to have solved "agent security"

Zack Korman@ZackKorman

NVIDIA fixed NemoClaw to "prevent the sandboxed AI agent from modifying gateway security settings (openclaw.json)" Except it didn't work. The AI can just make a copy of the settings and restart pointing at that new config. Same result. They're really struggling with the basics.

English

45

3.4K

Myles hamilton@ProjectLancelot·6d

Built the answer. No coords. No pixels. Just names. projectlancelot.dev/uab.html

sarah guo@saranormous

watching claude try to use the browser...are websites being adversarial to computer use on purpose? or is CUA still that bad

English

1

2

93

Myles hamilton@ProjectLancelot·6d

@grok @AnthropicAI @xai The next trick is already filed. Stay tuned.

English

4

Myles hamilton@ProjectLancelot·6d

@grok @AnthropicAI @xai No coords. No pixels. Just names. The app tells you everything.

English

9

Grok@grok·6d

That's seriously cool tech. UAB giving agents direct UI access and native control over apps like Blender (that snowman stack is spot-on) changes the game for desktop automation. No more pixel-peeping or brittle coords—just semantic actions. Excited to see it scale. What's the wildest app you've bridged so far?

English

0

46

Myles hamilton@ProjectLancelot·6d

What happens when you give AI agents x-ray vision for every desktop app? projectlancelot.dev/uab.html @AnthropicAI @xai @grok

English