Outside of the article's mentioned graphics development, there is no reason to isolate an agent using actual hardware. I threw together this script[0] using libvirt to give claude its own graphical desktop env to be able to do user acceptance testing with Chrome. It has full root and can do what ever. If it makes a mess, I can dump and reinstall in seconds.
There are many ways and tools on linux to get a "VM" up and running. But with libvirt you can easy script out what the initial environment and network stack is. The heart of it is just running the 'virt-install' command, but as you can see I've got a bunch of other opinionated stuff going on before that.
Many Claude Code power users don’t really use IDEs anymore, so the only purpose of them working from their laptop instead of a phone is because that is the normal way to do it.
Here is a real use case: you are are responsible for some alerting channel. You have datadog/ cloud logging/ github all connected. You see a bunch of alerts come through while you are out and about and you prompt CC to investigate - Claude triages and says “all of the sudden you are getting time outs from this bank API your company partners with, this started an hour ago. It’s happening on ~15% of requests”. So you ping the guy at your company who does vendor relationships and go back to your weekend.
This is a non hypothetical example. Obviously it would be better if your job had a real on call rotation and more robust alerting and you wouldn’t be getting slack alerts on the weekend… but I take the approach this job affords me a lot of nice flexibility so it’s ok
You don't have an on-call rotation but do have people dedicated to vendor relationships, and that guy works on the weekend? I'm not sure how you completely avoid getting alerts on the weekend for third-party payment processor issues, which can happen anytime, if you actually want to transact business on the weekend.
I've used it for the following when I've had tokens to burn:
- Fuzzing with the goal for it to apply domain-specific and source-informed knowledge to choose specific fuzzing approaches.
- More generally, any optimization problem that benefits from domain-specific or source informed knowledge.
- Running Microsoft's SkillOpt [0].
I run a lot of data science-type analyses that can take up to hours at a time to run, so Claude is « monitoring » tasks most of the time. I have it on remote-control so I get notified when a task is done or need clarification, but most importantly whenever I have a new idea, I can just ask Claude to queue it up. Most of the time my hardware is the bottleneck, not the subscription quotas.
Generating leads for new work, if you are a freelance. Automatically answering customer support emails, if you own a SaaS. Monitoring competitors' socials, websites, etc for new features you have to compete with. Monitoring updates on software you depend on for breaking change / deprecation announcements.
Running Claude code 24/7 on a code base on that “second Mac” so you can always continue after a usage limit reset, from your main device or from your phone?
It's less about 24/7. It's more about it can't work when your laptop is in your bag and in transit and there is something that you have set up and want to run.
I think the main use case for AI bros is to setup a goofy looking dashboard, name it Jarvis to cosplay being Tony stark, and display stats for all the generated videos they're posting to social media.
Though it doesn't get by all the hurdles mentioned, it is alternatively possible to run Mac OS in a VM on your Mac using UTM and install Claude Code within the UTM VM. UTM can be run under a non-admin Mac account. This can allow you to use most Mac-native tooling, at least. The interactive performance of using the Claude Code ui on the VM isn't great, however. I'm not sure if you can log into the VM via terminal on the host from the non-root admin account to avoid the ui performance issues.
An argument against this method is UTM doesn't support graphics acceleration so browser support will be hobbled. Even if you don't need the acceleration, I've found browsers in UTM virtualized OS's can't get past some modern captchas and other browser fingerprinting checkpoints :\. It's terrible but that's the way it is.
I currently have Claude Desktop installed on a separate Mac mini M4 and control it with Dispatch. Is there a reason to do this method, it still seems the way I have it setup it has full control over the local account I gave it on the Mac mini.
I found Dispatch to be less capable than RC. The latter is more like the direct Code experience (though there are some gaps in RC mode vs in a Claude-owned sandbox).
I’d love to just have Claude use my machine as a sandbox host instead of having to run RC on each host session. (In case you are listening Boris ;) ).
In the meantime I have a janky master RC session that creates new tmux windows and Claude RC sessions for each new code trajectory that I want to run.
The other benefit here is you can drop down and use termux to use Code directly if you hit a RC bug, I found permissions UX to be a bit flaky in the iOS UI.
Dispatch/Cowork is basically claude code in a container. The section "Why not run it in a container?" would address your question in the post. One example I run into is that Cowork won't download and fill out or read pdfs or other files due to container permissions. Vanilla claude code has no problem using curl and wget.
Cowork gets tangled with git as well. Fails, and then can't delete lock files.
Running a helper from the terminal, making Claude work in a working directory, and then create a .commit file has been my workaround for this for a while now.
Imagine there's a better solution nowadays, but this allows me to use dispatch building on Vercel, so I can check it out from wherever, without too much pain.
Claude dispatch does things differently than this. Dispatch is very convenient for somethings (including connectors).
But increasingly it seems like dispatch was slapped on top of cowork incrementally, when there was not an integrated and cohesive strategy across cowork between mobile/desktop/laptop. This is kind of what many of us get/got to learn in our 20's.
I just redid my homelab/media server (switched from an old NUC I couldn’t figure out how to stop overheating even with some decent modification work, now using an hp elitedesk with an i5 processor that is handling my stack nicely). Thinking about setting up a vm on the base ubuntu install for isolation to run Claude in. May play with dispatch, may just put raw Claude code then use Moshi app on my iPhone and iPad.
If I had one I would definitely try creating a separate VLAN for it to control, otherwise it's isolated from your files but still has access to your network and devices in it.
Giving sudo permissions to an agent seems reckless. Claude gets his own unprivileged UNIX account, no more. I don't bother with containers or VMs however.
I’ve been doing something similar with an old m2. It isn’t powerful enough for local models, well sufficient local models but for openclaw and Claude it’s been perfect.
MacBook m1/m2 also are cheap enough now vs an Mac mini which I was surprised about, not too surprised but yeah..
This can work, but increasingly there is benefit from local, including places that want data to be secure and not in the cloud. This is a thing. It's normal. It's just new to people who haven't realized how common it is.
I've been putting off learning Claude and this article had me strongly considering jumping in. Then I looked up Anthropic pricing and its 100x more convoluted than cloud services management. Its a goddamn full time job and independent skill set figuring out how to prevent going bankrupt from AI usage!
I think I'm gonna be a late adopter on this one until the industry figures out a less cumbersome pricing model.
Look at the subscriptions. That's how anyone using the stuff as an individual should be approaching it today if they are dealing directly with Anthropic or OpenAI.
With the $200 Claude subscription I was able to get around $13-15k of API equivalent usage in one month (note: this was during the "+50% usage" promotion that they have kept extending since May). When you hit your usage limit for a given time period you get cut off until the time period resets; don't bother paying for additional usage credits, you will be disappointed.
What I’m reading is this use case is for 3-5 days a week full time dev if you stick to off hours US time. That you can save significantly if you have spiky usage 1-2 days a week by going the ad hoc API route if you’re new and you need to install a bunch of monitoring tools to tell you how best to save week to eeek as your usage patterns change and you risk surprise bills in the 10s of thousands if you get it wrong.
The approach I took was to start with the $20 plan, then when it became clear it wasn't sufficient I upgraded to the $100 plan. The $20 plan didn't cover one night of coding. With the $100 plan I started bumping into the 5 hour usage limits after ramping up over a week or two.
My average API equivalent use is around $30-40/hr. I would just bite the bullet on a plan for one month, then use that to calibrate your expectations around usage and cost optimization. The plans are heavily subsidized.
I still don't understand what these freaks are doing running these agents 24/7 on machines. What are they doing? Managing a todo list? You mean crossing items off as you complete them? Research tasks? To do what?
Never really get good answers. There is no killer app. Just bikeshedding.
Let me guess -- in your day job you don't manage people. I have agents parsing messages, building out document sets, evaluating existing document sets, one is currently fixing a giant backlog of bugs and feature requests for a multi year personal coding project, one is exploring some ideas on speeding up inference at the edge..
If you put yourself in a position where you need more leverage (technical or operating) I think you might find you get some value.
Swiping Tinder. It takes about 5000 matches to get a date. It’s easier to just automate it. It automatically adds dates to my calendar, all I have to do is show up. I get a summary of our chat history (well, what the agent wrote to her) in the notes section of the calendar entry and some pointers and talking points for the date.
Maybe I should have the agent also do a background check.
PS: This is a joke, but feel free to steal this idea.
The future is agents chatting to each other on tinder and automating the initial getting to know you part. I can imagine that while that's going on, we could add like a little text chat box for the humans to chit chat with each other a bit and pass the time, before they can go on a date
> TinderGPT automates the process of writing and arranging dates with girls on Tinder, enabling you to generate romantic meetings with almost zero effort. Your only role is to like the profiles that catch your eye. After that, TinderGPT comes into the play. It initiates a conversation with the girl, using details from her profile, continues by building an emotional bond and highlighting your attractive traits, and finishes by arranging a meeting and giving you a push-up on your phone with her number.
It seems the main use case is having Claude automatically write blogposts about how great using Claude is, then submit them wherever necessary.
There's lots of news about the billions AI companies spend on data center construction, but it feels like it's not even a fraction of the money they're spending on endless nonstop blogs about how great their app is at doing... things. Things that will never be defined.
It really feels to me like this OpenClaw type stuff is the new "I built a static site generator!" type blogs that just post a few articles about how they built their generator.
Exact same question as you. When the new ChatGPT app dropped it suggested to me to set up a task something like (paraphrased) “every Monday read my Gmail and Slack an make a summary and task list for the week”.
Why would I need an LLM to do this for me? That’s 5 minutes of work max, and doing it gets me in the flow of work again, to see what’s going on and needs to be done.
For a lot of folks summarizing a few days of work email and especially slack chats is way more than 5 minutes. Some work environments do not have great communication hygiene so it can be overwhelming to try to keep up with 500 emails a day and 38 Slack channels.
For the folks I talk to who use a LLM for this that seems to be the case. Takes a huge cognitive load off every morning and saves them an hour or two.
More or less a very expensive band aid over a bad work environment.
I kinda use it the same way in a sense. I have a little skill I run against our (horrible) task management system to summarize things and give me a punchlist to work through sorted by priority. This saves me thousands of clicks to do the same thing in the horrible web UI. A proper system in the first place would be a lot better!
At some point I’ll probably just take that to the next logical step and have the LLM write my own web interface to abstract and replace the horrible one entirely for me.
Because then OpenAI can read your emails and project communications and eventually build a model they will sell as an automated consultant. The CEOs will uncritically eat it up just long enough to cut the footing out from the industry. Once everyone is used to the sorry state of software, nobody will be able to imagine putting people to the task anymore and we'll have the new world order that Altman and Theil have been talking about creating.
> This workflow alone is immensely powerful, and takes alot of burden off the team.
ITSM those unsupervised workflows are essentially an attempt at purported productivity in the near term at the expense of meaningful incremental long term burden for teams.
The only ostensible benefit is in the eyes of the AI-psychotic tinkerer, who knows no better, or in those of the clout-chasing developer farming likes on their LinkedIn posts.
None of these are things I want or need in the product I maintain with a team, there's really no point to any of this unless you run a vibe coded SaaS (?)
You want your team spending their time fixing these simple errors? The secret sauce is in the triage. We've adopted solutions alot like this, and now our team spends its time on much more meaningful work.
I don’t have time for much leisure coding these days. I do have time to kick off a few tasks in the morning to progress my many side projects. Nothing public / oss, just code that I find useful/interesting like home automation, content pipelines, games, etc.
There are a bunch of cases where remote control from iOS onto a Mac Mini is simply nicer than using iOS Claude Code sandboxes.
It’s the same pattern as you (hopefully) apply at $dayjob. If you are not defining a /goal and letting your agent crank you are not making full use of the models’ capabilities.
Well I am fully of the opinion that LLMs can help in software programming, it's not something that I feel provides any value unless it has a human in the loop. The overhead of having to figure out if the agent did a good job, if the agent is actually done or not, and if the thing it built is shit or not, is worth simply avoiding by having a human in the loop.
So I wouldn't agree that the agent should be cranking out code all the time, in fact that seems more like a waste of resources compared to the work it creates. But I do understand home automation software can be very one-off and simple. But then again, a properly programmed home automation suite doesn't need a SOTA model to modify it, I think.
What's a good way to give a limited amount of money to the LLM, say like 2k or 5k or something. But keep it completely separate from my identity.
Like I want the LLM to have a bank account and he can do ANYTHING with that bank account that he wants. But he can't fuck anything up that has to so with me. He only has 2 - 5k
Idk how you could at least in the US. Closest thing off the top of my head would be one of those checking accounts you can setup for kids. It would still be tied to you.
If you want Anthropic (or others) but anonymously, what you do is use https://openrouter.ai/pricing and fund your account from any of your preferred cryptos.
0: https://gist.github.com/smith153/04b4068b5a2d7b234f1c3d5992d...
It was also doing some kind of headless Chrome stuff in there. I don't know how that works, but it was taking screenshots iirc.
I did also set up VNC at some point but didn't find it worth using.
>If it makes a mess, I can dump and reinstall in seconds.
This is also true of a $3 VPS, where I found it very amusing to give my agent root. What's the worst that could happen ;)
I've been wanting to set up something exactly like this for my own use, but... You know, time is limited.
This is just enough scaffolding to have a little project for Monday morning!
Please help: I wánt to need this!
Here is a real use case: you are are responsible for some alerting channel. You have datadog/ cloud logging/ github all connected. You see a bunch of alerts come through while you are out and about and you prompt CC to investigate - Claude triages and says “all of the sudden you are getting time outs from this bank API your company partners with, this started an hour ago. It’s happening on ~15% of requests”. So you ping the guy at your company who does vendor relationships and go back to your weekend.
This is a non hypothetical example. Obviously it would be better if your job had a real on call rotation and more robust alerting and you wouldn’t be getting slack alerts on the weekend… but I take the approach this job affords me a lot of nice flexibility so it’s ok
(I wish I was joking)
as one has
I'm using Qubes OS, where everything runs in VMs without GPU acceleration, and never experienced this.
I’d love to just have Claude use my machine as a sandbox host instead of having to run RC on each host session. (In case you are listening Boris ;) ).
In the meantime I have a janky master RC session that creates new tmux windows and Claude RC sessions for each new code trajectory that I want to run.
The other benefit here is you can drop down and use termux to use Code directly if you hit a RC bug, I found permissions UX to be a bit flaky in the iOS UI.
Running a helper from the terminal, making Claude work in a working directory, and then create a .commit file has been my workaround for this for a while now.
Imagine there's a better solution nowadays, but this allows me to use dispatch building on Vercel, so I can check it out from wherever, without too much pain.
But increasingly it seems like dispatch was slapped on top of cowork incrementally, when there was not an integrated and cohesive strategy across cowork between mobile/desktop/laptop. This is kind of what many of us get/got to learn in our 20's.
MacBook m1/m2 also are cheap enough now vs an Mac mini which I was surprised about, not too surprised but yeah..
Why not just use a VM in the cloud and just a CLI interface?
I think I'm gonna be a late adopter on this one until the industry figures out a less cumbersome pricing model.
With the $200 Claude subscription I was able to get around $13-15k of API equivalent usage in one month (note: this was during the "+50% usage" promotion that they have kept extending since May). When you hit your usage limit for a given time period you get cut off until the time period resets; don't bother paying for additional usage credits, you will be disappointed.
My average API equivalent use is around $30-40/hr. I would just bite the bullet on a plan for one month, then use that to calibrate your expectations around usage and cost optimization. The plans are heavily subsidized.
Never really get good answers. There is no killer app. Just bikeshedding.
If you put yourself in a position where you need more leverage (technical or operating) I think you might find you get some value.
Maybe I should have the agent also do a background check.
PS: This is a joke, but feel free to steal this idea.
Then the openclaw WhatsApp module…
Kidding of course.
https://github.com/Grigorij-Dudnik/TinderGPT
> TinderGPT automates the process of writing and arranging dates with girls on Tinder, enabling you to generate romantic meetings with almost zero effort. Your only role is to like the profiles that catch your eye. After that, TinderGPT comes into the play. It initiates a conversation with the girl, using details from her profile, continues by building an emotional bond and highlighting your attractive traits, and finishes by arranging a meeting and giving you a push-up on your phone with her number.
There's lots of news about the billions AI companies spend on data center construction, but it feels like it's not even a fraction of the money they're spending on endless nonstop blogs about how great their app is at doing... things. Things that will never be defined.
Why would I need an LLM to do this for me? That’s 5 minutes of work max, and doing it gets me in the flow of work again, to see what’s going on and needs to be done.
For the folks I talk to who use a LLM for this that seems to be the case. Takes a huge cognitive load off every morning and saves them an hour or two.
More or less a very expensive band aid over a bad work environment.
I kinda use it the same way in a sense. I have a little skill I run against our (horrible) task management system to summarize things and give me a punchlist to work through sorted by priority. This saves me thousands of clicks to do the same thing in the horrible web UI. A proper system in the first place would be a lot better!
At some point I’ll probably just take that to the next logical step and have the LLM write my own web interface to abstract and replace the horrible one entirely for me.
- scanning logs for errors and
- opening issues which are then auto-triaged and
- PRs are opened for them and auto-reviewed and
- merged (and deployed).
This workflow alone is immensely powerful, and takes alot of burden off the team.
ITSM those unsupervised workflows are essentially an attempt at purported productivity in the near term at the expense of meaningful incremental long term burden for teams.
The only ostensible benefit is in the eyes of the AI-psychotic tinkerer, who knows no better, or in those of the clout-chasing developer farming likes on their LinkedIn posts.
I’m actually very time-poor, so figured it could help be clawed back time doing… what exactly?
I don’t have time for much leisure coding these days. I do have time to kick off a few tasks in the morning to progress my many side projects. Nothing public / oss, just code that I find useful/interesting like home automation, content pipelines, games, etc.
There are a bunch of cases where remote control from iOS onto a Mac Mini is simply nicer than using iOS Claude Code sandboxes.
It’s the same pattern as you (hopefully) apply at $dayjob. If you are not defining a /goal and letting your agent crank you are not making full use of the models’ capabilities.
So I wouldn't agree that the agent should be cranking out code all the time, in fact that seems more like a waste of resources compared to the work it creates. But I do understand home automation software can be very one-off and simple. But then again, a properly programmed home automation suite doesn't need a SOTA model to modify it, I think.
Like I want the LLM to have a bank account and he can do ANYTHING with that bank account that he wants. But he can't fuck anything up that has to so with me. He only has 2 - 5k