"DenchClaw finds your Chrome Profile and copies it fully into its own, so you won’t have to log in into all your websites again. DenchClaw sees what you see, does what you do."
Prompt injection as a service.
OpenClaw feels like early React" is a great framing but I think it leads to the wrong conclusion. React didn't win because Next.js showed up. React won because the component model was right, and then people built whatever they needed on top of it.
The risk with "framework on top of framework" is you inherit both set s of opinions AND both sets of bugs.. When OpenClaw updates, does DenchClaw break? When DuchDB has a quirk, is that your bug or theirs? The dependency chain gets real deep real fast.
Copying someone's entire Chrome profile is bold. Like, really bold. That's a massive ask. "DenchClaw sees what you see" is either the most powerful pitch or the scariest pitch, depends on who you talking to.
Well, of course I will test this thing you built in 2 days[1] for you!
[1]: https://xcancel.com/kumareth/status/2023534527113818625
Love the local-first approach. The "just ask it to import my Notion" thing via browser automation is really nice.
One thing I keep coming back to though - what if the tool could actually watch how you use your CRM and then suggest automations based on what it sees you doing repeatedly?
I've been building something called MemoryLane (https://github.com/deusXmachina-dev/memorylane) that does exactly this - it captures screen activity, spots repeated workflows, and suggests automations. Works as an MCP server so you can plug it into Claude or Cursor. Instead of you having to describe what you want automated, it just watches and proposes stuff.
Have you thought about adding something like pattern detection to denchclaw? Feels like it'd fit really well with the "everything app" direction. For us the most useful engine for executing skills and automations is surprisingly cowork thus far, haha
I get that trying to integrate AI tools into cloud systems can feel tedious and expensive. It’s frustrating when you want quick results but hit roadblocks with setup costs. Consider simplifying your stack by focusing on essential workflows to save both time and budget.
The local-first angle is interesting, especially for CRM data. Seeing this trend in observability and data engineering use cases as well.
cool
Great looking website! Looks like a genuinely useful implementation
This is a good approach!
The claw cesspool boldly thinks they are smart for just making all these things and probably thinks they came up with novel ideas, when everyone who has the slightest clue what is going on is petrified. Its clear these concepts are going to happen at one point, but we don't even have an answer how to do half of this safely. The worst part is that clawcels will use these for “outreach” and “content”
[dead]
[flagged]
[dead]
[flagged]
[dead]
Building anything on top of open claw seems wild to me. I checked it out due to all the hype and it was absolute dumpster fire of a project.
Lots of security vulnerabilities, but it was vibe coded. As much as I am an Anthropic Fanboy- I hope OpenAI can add some level of security to OpenClaw and they can get some good contributions. Although, these issues may be deep to its core.
Combining OpenClaw with sensitive personal data is a recipe for disaster.
... or disastrous comedy?
Interesting approach using DuckDB for the underlying CRM storage. I’m curious how well the file-system based structure scales once the dataset grows significantly.
CRM datasets don't really get that big. So DuckDB sounds like a great choice.
Seems interesting, but I see it's a fork of Openclaw that's many commits behind. Do you think you'll be able to keep DenchClaw updated with Openclaw?
I think a better solution would be to bring in one of the many Openclaw alternatives like NullClaw, ZeroClaw, etc. The magic of Openclaw is the heartbeat and cron modules so bringing in that piece should not be too difficult? I'll fork and hack away at it as well but the less dependent you are on other projects, the longer the longevity.
We started a direct OpenClaw fork, but we didn't want to always push upstream updates manually. So now we are completely detached from it. The GitHub says its a fork because it was originally forked, but now we have completely separated ourselves from it.
Now instead of bundling and patching from inside it, we just ship alongside OpenClaw so you can use the latest OpenClaw CLI separately yourself.
the crm isnt the hard part, the hard part is that most sales teams have a workflow problem disguised as a tooling problem. local first is smart but id focus on opinionated defaults for pipeline stages because thats where 90% of founders building their own crm get stuck, they model their process wrong then blame the software.
> created 17 minutes ago
Is this a bot lol, use words not buzzwords
Why would you build this on top of OpenClaw? Like, an insanely bad decision.
Vibe coded slop on top of vide coded slop to spam people? What could possibly go wrong?
exactly my thoughts!
I just use plain-text files for my CRM in Obsidian [1]. Works great if you are a solo founder only.
[1] https://www.ssp.sh/brain/managing-my-business-with-obsidian/
That's a simple but useful set up, thanks for sharing.
Love this setup! I also use Obsidian, but after DenchClaw I usually just open my Obsidian directory into DenchClaw so I can do anything with it. It has all the needed primitives for me like the markdown editor, graphs, etc.
Nice, this seems interesting. I don't use Obsidian (I use Logseq) but this has given me a couple of ideas for a CRM I am building (it's currently in a Personal Relationship manager phase which I've found useful for about a year or two).
Thanks for sharing.
Not a biggie, but might want to update the reference to 'Ironclaw' in the Try Ironclaw link at the top of dench.com
Oh yea
“Thanks for the heads up!” is what I would have said…
Are people using bottomless VC money to fund the API calls for these *Claw things?
I first tried OpenClaw with a local model, it gave poor performance. Then I tried it with Claude - great but it blew through hundreds of dollars in tokens in about an hour.
Or is everybody a billionaire now?
Lots of folks use the Claude Max 20x $200/mo. tier, I think. Gets you waaay farther than $200 of API credits.
No they are not. Claude Max does not give you access to APIs
The Claude subscription is only for Claude Code. Anthropic does not want people using the subscriptions with other clients.
Great, thanks for making me Google what CRM means in this context. Neither your post nor your website explains the acronym.
Sorry! It's basically a database for normies.
Customer Relationship Management.
Broad term for tools used to manage interactions with existing customers and/or sales prospects.
This is a pretty widely known acronym
I get the value of a personal CRM and potential power of having one locally managed by LLMs and I'd love to see such a solution, because to your point, outreach is just a small part of what you can do with a personal CRM. But, the way you describe and deliver this project is very confusing to me, it's a CRM but also Cursor for your Mac (what does that even mean?), I already run Cursor on my Mac, it also has a file tree view to use it as a better MacOS find I guess?
I think that a much cleaner messaging on what this tool is for would help.
Also a question about the implementation, why DuckDB for a CRM?
Something like SQLite feels like a much natural fit for a CRM where you primarily create, update and maybe delete records and you really care for the integrity of the data model.
From a quick look on the data model, everything seems to be a VARCHAR, if this is the case, why not just store everything in the file system instead? You do that with the md files and whatever is getting extracted from the SaaS tools.
I'm definitely biased here, but the OpenClaw hype is making people disregard the economics of it all. Building Auto-CRM.com, my primary concern was building a system that runs well while not costing 200$ per month to keep up, and of course, while also maintaining security. I assume the good guys at Folk, Pipedrive, etc also had similarly priorities. A lot of good work is being done within the OpenClaw ecosystem regarding RAG and memory, but specialised orchestration process to be a more reliable system.
I really want a DeathClaw product.
There is a Dungeon Clawer.
[flagged]
As of March 9 2026, it does log you out of your Google once on the main profile, but once you log back in everything works perfectly fine.
Question: Why do people want to do this with their mac mini? Can you not do all of this with a hosted VM instance? A mic mini makes it easier for people to set up? Everything still has to talk to data on the cloud right?
You can do it serverless, github and api keys is all you need - https://github.com/stephengpope/thepopebot
The openclaw-like system built using 'free compute' from github
Short answer: cargo culting. Mac mini is part of the ritual.
It's so they can talk to their bot using Apple iMessage. That's pretty much it.
Oh so you need an iPhone _and_ a Mac mini :)
Mac Mini is a gateway to iMessage. That’s really it.
Ah, that makes a lot of sense!
(Do not use imessage, a Whatsapp user, and we can access that through the browser, which means you can plug it into an extension)
The vast majority of americans have never used whatsapp, if they’ve even heard of it.
One on hand, this is genuinely cool. On the other end, this is the final nail in cold outreach's coffin.
Ha, I get why it looks that way from the CRM angle, but outreach is maybe 5% of what I actually use DenchClaw for day to day.
Yesterday I asked it to pull up all my meeting notes from last week, cross-reference them with my task list, and draft follow-ups. Before that I had it reorganize a messy folder of research PDFs into a structured workspace. I use it to build slides, write code (DenchClaw literally built DenchClaw), manage my calendar, search through old Notion pages I forgot existed.
The CRM part gets attention because that's what people asked for when we talked to power users. But the actual product is just "OpenClaw with a good UI, a file system, and DuckDB, running locally on your Mac." It does whatever you'd normally do on your computer. The browser is yours, the files are yours, the data never leaves your machine.
Think of it less as a sales tool and more as what happens when your entire Mac becomes programmable through natural language. The CRM is one app that runs on top of that. People are already using it for project management, research, personal knowledge bases, all kinds of stuff we didn't plan for.
I think this explanation is much easier to grasp :). And simpler to introduce to people, as many of the tech people still prefer better UX.
It's an advertisement mate.
Everything is skills. In a file system. That is the future.
Responding to some HN comments, I understand the focus on Sales Automation and Outreach can be worrysome.
But for me personally, this is where I do all knowledge work. For me it acts like Cursor, Happenstance, News Aggregator, Fun games creator like Pacman (it has an App Store), I can import Notion into editable MD files, create reports and presentations, etc.
That's exactly the direction I took with Filestash (https://github.com/mickael-kerjean/filestash) where everything is treated as a filesystem with fine-grained control to handle authorisation, plus a virtual filesystem layer to create completely new filesystems that don't 1 to 1 map to reality.
I've been meaning to use this in some project, you added a lot since last time I saw your repo!
In terms of "[XYZ] for agents", I think CRM is a big one that people haven't talked about as much. It becomes super relevant as soon as people start using an agent for anything customer related.
And the design principals are already pretty well established (accounts, contacts, leads, opportunities, custom object model, stages, etc.). It just needs to be turned into a database boilerplate with a bunch of agent tools. Excited to try this out.
Thank you, I'll be here for everyone to try it out, let me know how it goes!
Eventually there will just database tables, some skill files, and an agent
> It has a CRM focus because we asked a couple dozen hard-core OpenClaw users "what do you actually do", and it was sales automation, lead enrichment, biz dev, creating slides, linkedin outreach, email/notion/calendar stuff, and it's always painful to set up.
So basic automation and forcing the web to be "open"...
No one is talking about how AI is going to destroy business models that are dependent on dark patterns, on walled gardens, on poorly designed one size fits all implementations (so many things wedged sideways into sales force).
Yea, it has been a little shocking to me that the rising narratives around "AI agents everywhere" and "enable the web for AI agents" requires what we've all been wanting for awhile on the web (openness and interoperability) but that the same big players in tech have been clearly against for a long time. Like the fact that Google recently released that Google Workspace CLI (https://github.com/googleworkspace/cli) is a perfect example.
They could've released something like that years ago (the discovery service it's built on has existed for over a decade) but creating a simple, accessible, unified CLI for general integration apparently wasn't worth it until agents became the hot thing.
I wonder when / if there will be a rug pull on all of this. Because I really don't see what the long-term incentives are for incumbent tech platforms to make it easy for automated systems to essentially pull users away from the actual platform. I guess they're focused on the short term incentives. And once they decide the party's over, promising upstarts and competition can get absorbed and it'll be business as usual. Idk, we'll see.
FYI from your project link:
Note
This is not an officially supported Google product.
Can my agents (powered by NanoClaw or Claude Code) use the CRM without installing OpenClaw codebase?
This is an OpenClaw framework, so it installs / relies on your existing OpenClaw codebase. I think there has been a ton of requests on Claude Code support, someone has been working on a PR for exactly this, I'll update you here if it ships.
Readme has a discord link claiming 25K online, might want to update that it's quite misleading.
Sorry, it happened because it was originally an OpenClaw fork.
Are small local models good enough for driving OpenClaw-likes or an API key from one of the big labs is needed?
Depends on what you're using it for I suppose. A common tactic with Openclaw itself is to have a cheap or local model as the default, with rules to "escalate" to other models based on task complexity/type. But if every cron job comes with complete access to your personal machine and browser profile... Yeah, better go for the most predictable model you can find.
I always recommend Claude Opus 4.6 for anything OpenClaw gets its hands on.
I was hopeful that small models would be enough for the majority of the tasks, but I understand that it's not the case yet. But thanks for answering
> It has a CRM focus because we asked a couple dozen hard-core OpenClaw users "what do you actually do", and it was sales automation, lead enrichment, biz dev, creating slides, linkedin outreach, email/notion/calendar stuff, and it's always painful to set up.
Fuck me, it's going to get worse before it gets better, isn't it?
I've taken that bit out of the text above - I originally advised Kumar to put it in there (it's actually from the opening of the demo video), but in hindsight, I should have known it would backfire with the HN audience.
100% :-/
In response maybe we should design TCPAclaw. It is specialized in honeypotting all of the random cold call spam, tracks down the source of unsolicited contacts; including registration state, legal contacts, and registered agent(s). It then drafts and sends a TCPA letter and waits for one of two things to happen: Either a $500-$1500 check arriving in your mailbox, or the demand deadline elapses. In case of demand deadline elapse, TCPAclaw files a small claims suit in the appropriate court of jurisdiction.
Fight fire with fire.
I'm in.
That's... not a bad idea. The downside is the bot would be doing a lot of these and false-positives would be... embarrassing (like a real investor outreach).
Am I the only one that read this as "DeathClaw"?
Sounds like a great name for a chaos-fork for Openclaw.
> sales automation, lead enrichment, biz dev, [...] linkedin outreach,
Sigh.
It also does all or most knowledge work there is, the goal is for it to be smartly be able to do anything you ever do on your machine.
I've taken that bit out of the text above. See https://news.ycombinator.com/item?id=47314105 for more.
Looking at that star graph: Since OpenClaw became a thing, I can't help but conclude that Github interest/popularity metrics have become useless signals.
> popularity metrics have become useless signals.
That can happen!
Especially considering this project is 2 days old and has 580 stars. 500 seems like it would be a nice round number if one were to purchase bot engagement. Not confident enough to make that claim directly, but something about this project doesn't sit right in general.
Bruh it's not botted, the 500 stars came from Garry Tan's viral tweet.
Would also be a good cover up...
Can you link to it? I'm not able to find it on his account. Unless you mean his retweet of your tweet? If so, that retweet has just under 10k views and the tweet is in celebration of hitting 500 stars on Github.
Sigh The HN post literally has the link... https://x.com/garrytan/status/2023518514120937672?s=20
Wow, sorry, but given how incredibly insecure all the "claw" agent type things are right now, does this really sound wise at all?
It sees everything you do, really? What's it gonna do with that data? You don't know.
Put all your customer data in there, all your customer relationships. It's fine, it couldn't leak all that information, it couldn't screw up any sensitive business details I'm sure. This is gonna go great.
Sorry AFK everybody I'm gonna go get myself a VibeMBA.
Anyway, good luck, I'm really looking forward to the user stories in a few weeks! I'm sure this won't go badly at all.
> DenchClaw finds your Chrome Profile and copies it fully into its own, so you won’t have to log in into all your websites again. DenchClaw sees what you see, does what you do. It’s an everything app, that sits locally on your mac.
Wow that sounds great. Hey don't worry these things never blackmail anyone. Let it know if you're gonna turn it off, I bet it'll make some REAL interesting choices based on your browsing history
I'm always confused by this kind of comment about AI accessing people's chrome history because it seems to imply that the kind of person who uses this tool is both too stupid to know what private browsing is and also is into absolutely heinous stuff.
I feel like the average person is going to be like "oh no it'd be terrible if everyone found out I really like the 'big boobs' category on pornhub"
Privacy and security and whatever this could trample all over are not the same thing.
You may be legally entirely above board (though Cardinal Richelieu wouldn't let that get in the way) but you still might not want your S&M kink to be known or to be outed to conservative friends and family or have your bank account details spread around or have a $$$$$ bill run up in your AWS or LLM logins...
Oh, you have nothing to hide? Kindly paste all your payment and login credentials that your browser stores. Later we'll need to see all your DMs on Facebook, LinkedIn, Slack, Discord, etc.
Finally we'll want to know about disputes you've had with intimate partners, employers and other service providers, especially powerful ones like healthcare, insurance and financial organisations.
We should also have full published salary and benefits (etc) details right now, whatever their contract says about disclosing those, and 24x7 streamed video of their entire life with no censoring, including toilet breaks and sex and bars and parties.
And, along with all the credentials as you suggest, including private parts of PGP keys etc, accurate impressions/clones of any and all physical security/privacy devices they use such as keys to house and car and safe and gun safe and relatives' crypt, etc, etc...
[flagged]
[dead]
Data gathering / creating / updating / filtering / creating reports, Doing certain action on every data entry (like sending email), etc.
Telling DenchClaw to "make it less robotic" on 300+ personalised drafts is still better than me actually making it less robotic myself imo
[flagged]
Dench, is that you?
I can't tell if this is a bot or human response.
It’s definitely coming across as having being written by an LLM
OpenClaw opens a wide attack surface on your digital life that cannot be remediated so long as hallucinations and prompt injection remain unsolved problems. Anything built on top of it is equally insecure and probably even more insecure.
I really don't want to yuck anybody's yums or step on dev work that I had nothing to do with, because I've been there and I know it sucks, but OpenClaw is barely secure enough to even play with in a sandbox. Giving it private information about your real business and real business contacts feels like an absolutely insane thing to do.
At best OpenClaw is like a toy... if the toy was a gun and it shot real bullets. This feels like playing Russian roulette with your livelihood.
I cannot fucking believe people are letting it remote start their cars and control their garage door. Nevermind ovens. All things people have done and posted about.
As someone that has worked in the automotive space, an enormous amount of regulation and effort is spent making sure you cannot do something like forgetfully remote start the car with your garage door closed and gas yourself. Nevermind securing it so that others cannot do this to you.
And these people are plugging it into ... this, which will happily go "oh, the car turned off after 15 minutes, let me turn it back on!"
There are realistic odds that someone is rotting in their house while their lobster pays the bills and writes blog posts for them.
This rings so true. Software Engineering should have stricter bar similar to med professionals. If we have leaked such lousy products and the public crowd thinks this is usable, it's a failure of the industry as a whole.
>> Software Engineering should have stricter bar similar to med professionals.
This is a month-old project by someone how has been suckling at the YC teat of release as early as possible; #YOLO. There's no "engineering" here.
> so long as hallucinations and prompt injection remain unsolved problems
Aren't hallucinations mathematically impossible to be _solved_? Cannot believe how so many people just willy nilly give everything they have to a lying parrot.
I like the idea of OpenClaw a lot, it's a technology that I would want in my life. But in it's current form it's kinda chilling and I cannot see it become safe to use anytime soon.
It seems to me many infosec best practices that have been built over decades have been forgot in the last few months like nothing happened. People really do give this kind of software full system access, plus access to their emails, their private chats, most likely their passwords too and who knows what else via plugins. I couldn't really imagine this happening one year ago.
I'm 100% confident that any state actor and cybercrime groups are currently heavily focusing their research on these tools. You compromise the right person and you can access all kind of critical information, it would basically be the same as having some remote control software on their system with full permissions.
And everyone on the hype train seems to be absolutely unaware of this. Maybe I'm missing something, but all of this feels so odd to me.
I think a lot of them are aware of it, but also grifters, and hoping to profit off of it before the bomb goes off so that they can claim ignorance and escape blame. New and powerful thing that people don't fully understand becomes fertile ground for grifters to sew their sins. Like when Marie Curie discovered radium and everyone and their mother started forcing it into products, including toothpaste and "medicine", within like 5-10 years.
At what point does this become an AI powered spamming machine?
[astronaut with gun meme] Neal Stephenson depicts this outcome in his novels as "The Miasma" and introduces a zero knowledge biometric based cryptography scheme used by everyone to validate content, and everyone has to have advanced AI filters in order to pluck out tiny tidbits of signal from among the noise.
We're going to need local AI to sift through the trash. Platforms have been more or less useless at curating content, and it's only smaller sites like HN that have retained a high SNR at this point. It doesn't even matter what media, at this point, video has passed the 2-3 second sniff test. We're seeing boomers get completely sniped by AI videos, even with watermark, showing absurd spin on current events. Text, music, podcasts, video, cartoons, whatever, it's all been infested, and the quality keeps increasing. I've seen a couple 2+ minute seedance productions that have been actually enjoyable, but by June that sort of thing will be one-shot prompting instead of someone gluing together the outputs from 4 difference SoTA AI tools.
It's getting weird, and we're not ready for it, at all.
Truly tired of seeing yet-another spam machine. All of these hype machines are built to spam people about their /paid/ hype product, rinse-repeat. BS
I get why it looks that way from the CRM angle, but outreach is maybe 5% of what I actually use DenchClaw for day to day.
Yesterday I asked it to pull up all my meeting notes from last week, cross-reference them with my task list, and draft follow-ups. Before that I had it reorganize a messy folder of research PDFs into a structured workspace. I use it to build slides, write code (DenchClaw literally built DenchClaw), manage my calendar, search through old Notion pages I forgot existed.
The CRM part gets attention because that's what people asked for when we talked to power users. But the actual product is just "OpenClaw with a good UI, a file system, and DuckDB, running locally on your Mac." It does whatever you'd normally do on your computer. The browser is yours, the files are yours, the data never leaves your machine.
Think of it less as a sales tool and more as what happens when your entire Mac becomes programmable through natural language. The CRM is one app that runs on top of that. People are already using it for project management, research, personal knowledge bases, all kinds of stuff we didn't plan for.
>but outreach is maybe 5% of what I actually use DenchClaw for day to day.
will you be enforcing the same for the users of your product?
if not, i am not sure how this statement addresses the above concerns.
How do you enforce this on an open source github project?
you cant, which was sort of my point
Become? I believe that’s the point.
Cold calling is not 'spam' because it is essentially done by a human. This is no different than an email spam network. So now this will just become email / linkedin spam done by corporations? I guess we turn up the filters now?
Just because a human gets paid to sit at a computer calling random people doesn't absolve them of a spam title.
I agree that it is spam of a sort, but I don't think that's how it's generally portrayed. If biz dev and sales are just spammers (because of LLM automation) then we should reclassify them and shun those types of posts.
I find it amusing that one of the main things to do with OpenClaw and other similar tools is create a Web Interface on top of it so that users can click on buttons when the entire promise of the technology is that you don't have to do any of that because it transcends standard UI.
I mean, ultimately why would you even need a CRM if not to sell something? And if you are going to sell something ultimately you want to get that done without any additional layers of abstraction. So the interface is the definition of the goal and the output is measured in results.
"Hey claw, I want to sell my product. Go figure it out!"
You don't need a UI for that.
To make this easier we could develop some kind of… query or… command language for this.
Take OPs example…
> I am addicted now that I can ask it, “hey in the companies table only show me the ones who have more than 5 employees”
Now how could that command language look like, maybe something like…
PICK * of COMPANIES if EMPLOYEE_COUNT >10;
We could call this DCCL: Dench Claw Command Language!
It is called SQL and the tool is called sqlite.
That’s the joke :)
[flagged]
The way imports work in DenchClaw is a bit unconventional, when you tell it to "import my HubSpot", the agent literally opens your browser (using the copied Chrome profile), navigates to HubSpot, triggers the export, and then ingests the downloaded files into the workspace DuckDB. So the bottleneck isn't really a fat in-memory ETL... it's more like processing a CSV/JSON export file on disk.
For the DuckDB side specifically: we shell out to the duckdb CLI binary for every query rather than embedding it in the Node process. So each operation gets its own memory space and dies when it's done. the web server at localhost:3100 stays lean regardless of what you're ingesting. DuckDB's out-of-core execution also means it can handle datasets larger than available RAM natively, which is one of the reasons we picked it over SQLite.
For really large exports (think full HubSpot instance with 100k+ contacts), the practical limit is more about the browser export step than DuckDB. HubSpot itself chunks its exports, and we process those chunks as they land. The DuckDB insert is the fast part.
Honestly for CRM-scale data, even a large sales org's full HubSpot, DuckDB eats it for breakfast. Where it would get interesting is if someone tries to throw analytics-scale data at it, but that's not really the use case. Would love to hear how IndexedDB holds up for you at scale in AccIQ, different trade-offs for sure.
> The way imports work in DenchClaw is a bit unconventional, when you tell it to "import my HubSpot", the agent literally opens your browser (using the copied Chrome profile), navigates to HubSpot, triggers the export, and then ingests the downloaded files into the workspace DuckDB.
What’s stopping the agent from doing literally any other thing in HubSpot? You know, small stuff like editing/deleting records, sensing emails, launching marketing campaigns, deleting reports, etc.
Our HubSpot import seed skills have strong always on prompts for asking user before doing any action, and it knowing where to click. For actions faster than browser, the skill also knows how to use hubspot cli.
Ideally for these pursposes, I would ALWAYS use Claude Opus 4.6 for this stuff, personally I have never seen it do unintended things to that extent.
Also, when the browser opens you can supervise it doing the thing, since you can see what its doing, you can always stop it if it ever goes wrong.
Right, but you and I both know that skill files are merely suggestions that the LLM often but not always follows