How Do I Know What Grok Bot Did With My Accounts?¶
Grok Bot is genuinely convenient these days, especially now that it has a password vault (Vault). Things I used to have to open a browser, sign into an account, and click through myself can now be handed over to it. Most of the time I just tell it what to do, and it gets it done. It saves me a huge amount of time.
But here's the problem.
Sometimes it gets stuck. Sometimes it clicks around on a website. How am I supposed to know what it actually clicked? It tells me the job is finished, but did it accidentally delete something else? Did it press the wrong button and never mention it?
With Vault, Grok Bot can also use accounts I've authorized to operate websites. That's wonderful for convenience. But trusting an AI completely with all of that? Honestly, it's a little scary.
And if there's no record afterward, I don't know how to trace what happened, let alone undo it. Give it enough time, and maybe even the bot itself won't remember what it did.
So I started wondering: what if Grok Bot recorded the entire process whenever it used a browser?
That's why I built Agent Recorder MCP.
Once Grok Bot has it set up, it can start recording before it uses the browser and save what happened on its own screen as an MP4. When the job is done, it sends me a link through FastFileLink. On an ordinary network it prefers P2P transfer; in some restricted environments it uses a relay instead.
The video is stored on the agent's machine first. Only after I finish downloading it and the transfer service confirms completion does the local recording get deleted. So it doesn't keep filling up the agent's disk. If I want to save a copy, I can also have it uploaded through a connected Google Drive account.
How do you get Grok Bot to install it?¶
I deliberately made setup as simple as possible. Ordinary users shouldn't have to learn a pile of MCP commands.
I wrote a Skill that tells Grok Bot how to set up the recorder, and when to start and stop recording.
All you have to do is send Grok Bot this link:
Then tell it:
Please read and install this Skill, and follow its instructions to set up the recording feature. Whenever you operate a browser for me, start recording first and give me the replay link and a summary when you're done. If you need any permission or confirmation from me, just ask. Finally, make one short test recording to confirm the video plays correctly.
That's it. Grok Bot follows the installation steps supported by its environment and asks you to approve anything that requires permission. If the environment doesn't support recording, it should tell you rather than pretend everything is installed.
Once setup is complete, have it open a simple webpage as a test and check whether you actually receive a playable video.
Do I have to do anything differently afterward?¶
No. I just keep asking Grok Bot to work the way I always have.
I might ask it to sign in to a website, check some settings, or handle the clicking I'd otherwise have to do myself. As long as it follows the Skill, it starts recording before touching the browser and stops when the job is finished.
At the end, I get more than a message saying the job is done. I get a replay link and an action summary. The summary roughly covers the steps it took, the systems it worked with, any problems it hit, and how it dealt with them.
If I want the details, I can open the video. No need to keep asking, “Wait, which button did you click?”
Won't all these recordings take up space?¶
Not permanently, by default.
The MP4 starts out on the agent's own machine and is shared through FastFileLink. Once the download completes, that local file is removed. The video isn't fed into the AI model, and it isn't deleted just because a link was created.
I can also choose to have it saved to Google Drive. In that case, the agent's connected Google Drive tool handles the upload.
One important limitation: right now, this records the agent's own Linux X11 desktop, not my Windows or Mac screen. It's not an operating-system-level audit system that forces every action to be captured. Grok Bot still has to follow the Skill and start recording before browser work if I want the whole session on video.
The recording might also include private data, admin pages, or even private keys and verification details. That's something to be careful about when sharing or storing the video.
I feel a lot better about letting it do things¶
For example, I later had Grok Bot help me with the process of requesting an Apple private key. Having the recording made me much more comfortable.
When it finished, it gave me a rough explanation of what it had done, and I could also watch the video to see how it had actually worked. If something looked wrong, at least I had something to go back to instead of wondering, “Did it delete something? Where did it go?”
Of course, a recording doesn't stop mistakes, and it doesn't guarantee that you can undo them. With private keys and sensitive accounts, you still need to check the important bits yourself.
But at least I no longer have to take “Done!” at face value with no idea what happened along the way.
I think this is something everyone who asks an AI to operate websites for them should have. AI saves us a lot of time. Having a recording gives me one more reason to feel at ease.
Give it a try.
Open-source project: company-notes/agent-recorder-mcp