There’s a new personal agent in town.

(Well, there are actually two. The mysterious Instinct just revealed itself as I was writing, and raised $250 million at a $2.5 billion valuation.)

It’s represented by a blob-like avatar that looks straight out of the Scream series.

And it’s run by a wild man who wants to colonize Mars. More importantly, by the team at Cursor that just knows how to ship.

I had my hesitations on giving them access to my sensitive information (more on that below).

But more importantly, I thought it was going to be another fad that we’d forget about in 2 weeks.

I was wrong. I’m using it all the time now. Here’s how I tested it.

1. How is the mobile experience?

But first, I’ve flirted with a lot of personal agents. And one of my key criteria is that the mobile experience has to be seamless. That’s particularly hard, since all the powerful tools (Cowork, Claude Code) need to run on your local machine.

It took a while to get OpenClaw running (and I loved it, until they shut down the Opus subsidy).

Today, I’m bouncing between Codex’s remote feature (and to a lesser extent Claude Code’s remote control).

And while they’re really good, they’re just not it.

They don’t pass what my friends at Every call the “reach test.” I use these tools because they’re the best available option. But I never reach for them, mostly because they have weak mobile experiences.

So to begin, Grok Bot has a very strong mobile experience. It’s basically a few chat bubbles with minimal extra features and things to configure.

Mobile Experience: ⭐️⭐️⭐️⭐️⭐️ (out of 5)

2. How do you access my context?

If you’re going to bounce around different harnesses, you better make adding your context easy. Given Grok’s track record of weak data security, I was careful about what sources I’d connect.

Plug-ins and Connectors

Grok’s plugins marketplace is pretty simple (and looks like all of its peers):

I ended up connecting Gmail, Calendar, Granola and Notion. I have 3 very detailed wikis inside Github and decided to pick the repo that only contained information from public sources.

Two of my tools didn’t have plugins (Readwise and Todoist), but Grok Bot easily connected to them via chat.

How about my other context?

This is where things got confusing.

I’m so used to working out of a local folder (both with Cowork and Codex) so when I didn’t have this option it felt like the rug was being pulled from under me.

I’ve spent so much time curating context in markdown files (by client, workflow and functional units) that I can’t imagine using AI without access to this rich context.

This made me immediately skeptical of how powerful Grok could be, since it was effectively relying on connectors as its primary context source.

Context Experience: ⭐️⭐️⭐️

(5 stars for plugin experience, 2 stars for lack of curation)

3. How do you access a computer?

MCP is not enough. It’s become crystal clear that AI needs access to a computer. Claude’s Chrome extension remains criminally underutilized and Codex has taken computer use to an entirely new level given its speed, accuracy and cost efficiency.

But this is the double-edged sword of local machines.

You love them because they can act like you (and do things like log into your LinkedIn or view your WhatsApp messages).

And you hate them because your computer always has to be on.

Grok Bot takes a totally novel approach here, by creating an actual computer in the cloud just for you. (Technically, called a virtual machine.)

Here’s me logging into my gym:

My initial reaction was that it was clunky. It was very hard to copy-paste passwords into this laggy screen. And a few sites (like Mindbody, for my gym) would repeatedly log out.

That being said, ultimately I don’t want my physical computer to be part of my AI infrastructure and I look forward to the day when everything runs in the cloud. This is a great first step.

Computer Use Experience: ⭐️⭐️⭐️⭐️

4. How resourceful are you?

One of the best parts of AI is its ability to connect dots. With just access to your email and calendar it could tell you that your mom’s birthday is in 2 weeks and you should consider getting her a gift. It would then look at prior interactions and try to infer what she might want. Then, it could go as far as looking up specific products on the web and adding them to your cart.

There’s a delicate line here between being resourceful and creating work for the sake of creating work.

Because Grok Bot doesn’t have a ton of context, I was curious to see how good it was at connecting the dots.

For example, when I asked:

“Can I be helpful to any prospects?”

It did the following:

Looked up the actual prospect names in Notion
Look up any related notes in Notion
Search through Granola and email for prior interactions
Make the recommendation

This is the kind of task that is best measured via vibes. And the vibes were good (again, resembling OpenClaw using Opus).

Resourcefulness: ⭐️⭐️⭐️⭐️

5. What frictions do you remove? (Or add?)

This is where I think Grok Bot shines. It removes a lot of frictions including:

One chat per agent
The connectors work seamlessly
Computer use works well (once configured)
It’s fast
There’s not too much to customize

Here were the frictions that were added:

I still am not really sure where skills, memory and other context live
I’m not sure if I have too many (or too few) chats

Friction Removal: ⭐️⭐️⭐️⭐️

The final verdict

Grok Bot currently passes the reach test for a large number of personal tasks.

It’s not the heavy hitter that I bring in to do deep client work/analyses or to build out tools/web apps.

While the lack of customization requires me to “trust it” more, when I do, it usually delivers.

I write the Future-Proof your Career with AI newsletter, read by 10,000+ professionals and entrepreneurs.

Grok Bot is surprisingly good

Job breakdowns