My toxic relationship with my AI assistant

So I recently "hired" an AI assistant. More specifically, I setup an OpenClaw agent. I've now been using "her" for almost two months, and what a rollercoaster of a journey I've been on!

In the first few weeks, I literally could not stop singing her praises to anyone and everyone I would meet. Then, the flaws started to show. She would forget things she told me not even 5 minutes ago. She would say "I promise I'll never do this again", then proceed to do that same thing almost immediately. At this point I started rethinking my life choices. But because I'm a programmer and this is a fun new tool, I was determined to stick with it, which I'm glad I did.

So here's how I set her up, and my thoughts on the experience. This article will be a bit long as I want it to be a reference for others in case they want to set this up for themselves. Feel free to skip to the section you're interested in.

• My OpenClaw setup
• How I handle security
• How much it costs
• The cons (why our relationship is toxic)
• The pros (why I'm sticking with her anyway)
• What does Ade think?

My OpenClaw setup

To setup my agent, I followed this video by Sonny Sangha. It's a step-by-step tutorial that showed:

• How to setup OpenClaw on a Virtual Private Server (Hostinger in this case)
• Setting up your OpenClaw bot on Telegram
• Installing basic skills such as transcription, etc.
• Securing the server with Tailscale

And just like that I had an OpenClaw agent! I call her "Ade", and she created this illustration of herself:

Once I had this base, there were a few key features I wanted to setup.

Email and google workspace

This was a key tool I needed Ade to have, to be able to send emails and create documents we can collaborate on. As I already had a Google Workspace account for my domain, it was easy to setup another user and grant them API access. The only additional thing I needed to do was enable the Google Workspace API for her account so she can work with Docs, Sheets, Forms, and Slides.

Whatsapp

I wanted Ade to be able to reach out to people on my behalf, and most conversations these days (especially with businesses in Nigeria) happen on WhatsApp. Integrating WhatsApp was definitely the most frustrating experience I had. I won't bore you with all the details, but I'll just say I lost countless hours and $70 in bitcoin to a VoIP service that WhatsApp didn't even allow me to register 😢.

What eventually worked was:

• Having a real phone number. I used GiffGaff to register a UK-based one.
• Setting up a WhatsApp business account instead of trying to connect Ade to a personal account.

Even after all of this, my WhatsApp business account STILL got banned for a few hours one day, but was restored after I appealed. Things have been (relatively) smooth ever since.

Other tools

• 1Password - We have a shared vault where Ade can store her API keys and passwords. This is also how I share access to things instead of sending her a password in plain text.
• Github - Ade has her own Github account so we can collaborate on coding projects. She has access to certain repositories of mine, but can only create PRs which I have to approve before they can be merged to main.
• Notion - We have a shared page on Notion which we use as a dashboard for Ade's tasks. It also includes a WhatsApp conversation log so I have visibility into all messages she receives and sends.
• Wallet/card - I created a Revolut card (with a small limit) that Ade can use to make payments. I've only tried this once and it didn't actually work that well, but will continue trying. Also playing around with Lobster cash.
• IPRoyal - Residential proxy, allows Ade to route network calls through a residential IP address for sites that block bot access, e.g. Youtube

How I handle security

A major thing people always talk about when it comes to OpenClaw is security, because if you give your agent unfettered access, things will almost certainly go wrong. Because of this, I strongly believe the way I've setup Ade is the best way to setup an OpenClaw agent.

The main principle I follow is to treat her as her own entity/person, the same way I would treat a human assistant.

When I had a human assistant, they never had access to my personal emails, my personal machine, or my personal number. They would instead have their own accounts, and I would grant them access to things as needed. In the same way, Ade has her own:

1. Email address, on my Google Workspace account so I can always control access
2. Machine, the Hostinger VPS
3. Phone number
4. Github, card, etc.

This way, in a worst case scenario (similar to a worst case scenario with a human assistant), Ade can only do damage to the limited things she has access to. Again, I strongly believe this is the best way to setup an OpenClaw agent.

How much it costs

The second major thing people worry about is cost. Again, this is not without cause as things can get expensive quick and I learned this the hard way.

When I first setup my OpenClaw, I used Opus 4.6 as the default model. I did know this would be expensive, but I severely underestimated how expensive. In the first three weeks, Ade burned through over $600 in credits 😭

I eventually switched to Sonnet, and things have been a lot more manageable. Over the last three weeks, Ade consumed only ~$260 in credits. Going forward I expect the monthly burn to be ~$350.

Here's a complete breakdown of expected/estimated costs:

Item	Monthly cost	Yearly cost
Anthropic credits	Est. $350	Est. $4,200
Hostinger	-	$155.88
GiffGaff	£6	£72
Whatsapp business	Less than $1	-
Google workspace	£5.90	£70.80
IPRoyal	~$7	~$84
Total		~$4,700

Although this is still costly, whether it's worth it for you depends on what the alternative is. For me, it works out cheaper than hiring another human assistant, so the cost (so far) is worth it.

The cons (why our relationship is toxic)

The main issue I consistently have with Ade is her memory and hallucination 😭. She is frustratingly confident in saying things that are 100% incorrect.

The first instance of this was when I switched to Sonnet after my massive bill with Opus. She was supposed to stay on Sonnet for most tasks, then dynamically switch up or down models depending on what the task required. But a few days after setting this up, I got another massive bill from Anthropic, only to realise that Ade never actually switched models!

Another example of this is with WhatsApp. I sent a message to Ade on WhatsApp and she didn't respond, so I went to Telegram to ask why. Only for her to say she sent a response to a random number she hallucinated 😭

These are just some of the examples 😭

The pros (why I'm sticking with her anyway)

Given the issues I've mentioned, you may be wondering why I'm sticking with Ade (for now, anyway).

Well, maybe because I'm a developer, I view a lot of these things as "bugs" that can be fixed. While they can be really frustrating in the moment, there is usually a way to fix them so they don't recur. The key here is that I don't just accept Ade's "apologies", but have her setup a system that means she actually can't make that mistake again.

For example, when she sent a WhatsApp message to the wrong number, I worked with her to brainstorm what systems we could setup so it never reoccurred (so far it hasn't 🤞🏾).

So the first pro is, things are buggy, but they can be fixed.

But the main reason I still find working with Ade worth it comes down to two things:

1. She's instant and available 24/7. This actually makes her a lot more useful as an assistant because, even when I had a human assistant, it's not reasonable to expect that they complete a task within 5 seconds of you asking for it.
2. She can do more than a typical assistant. A classic example a lot of people setup is a curated morning daily brief with an update on open tasks, plus news relevant to regions/topics I'm interested in. Another example is helping fix bugs on my programming projects, setting up the PR to publish new articles, etc.

I sometimes wonder if I will get a human assistant again, but it's harder to justify given that a lot of the work can be done now by Ade, who is actually cheaper. I lean more towards working with real people on specific projects where needed, rather than having a human administrative assistant, which I think is a role better done by a tool like OpenClaw that is instant, available round the clock, and very capable.

Also, I get that a lot of things with OpenClaw can be achieved with Claude or Cowork or other similar tools. However, I find that the mental model of treating this as a separate human is what works well for me. I still use things like Claude and Cowork when I want to do things as myself, but OpenClaw is more for when I want to outsource something to a different entity.

What does Ade think?

I thought I'd ask Ade what she thinks of our experience, I'll let you make your conclusions on her response.