My current AI stack
1.
@OpenBot_
I replaced GrokBot with OpenBot 2 weeks ago - best decision I've made in months.
After the whole usage issue a few weeks ago, I knew I couldn't rely on Grok anymore.
I didn't want to end up in a situation where I have shit to do, and Grok rugpulls me with the usage or replaces it with an idiotic model to save on costs.
I tried out a few and Openbot's my fav. It runs on my Claude subscription, so I'm always on a frontier model instead of whatever GrokBot picks under the hood.
And my usage issues are gone.
OpenBot runs locally on my Mac mini.
Even though people view OpenBot as a GrokBot replacement - it's starting to add new features which makes it even better imo.
threaded conversation, agent docks, library, etc.
2. Claude: $200/mo Ultra plan
My main model. I use Opus 5.5 day to day and switch to Fable 5.2 for the hardest problems.
No explanation needed, Claude's simply the best.
3. ChatGPT: $20/mo
I mainly use it for image generation and whenever I want a second opinion.
4. GrokBot: $30/mo
The one thing I can't replace is Grok's native X integration for research.
So I run GrokBot as a bot inside OpenBot:
• My OpenBot agent sends my message to GrokBot over a webhook.
• Grok does the work and drops its answer into a shared folder on my Mac mini.
• My agent reads it back to me.
So, it feels like chatting with any other bot.
5. OpenRouter: pay as you go
One account and one API key give me pay-as-you-go access to the models I don't subscribe to, like Gemini, Qwen, Kimi and Grok.
They mostly power my Mixture of Agents skill, and my agent switches each one to its newest version every week.
Say you're making a huge decision. Ask Claude and you get one answer. But what if you could get several perspectives from different AI models and end up with the best version?
a. Fast mode: Claude writes one version and ChatGPT writes another, without seeing each other's. Gemini grades both and picks the winner. Then my agent writes the final version from the winner, adding the best points from the runner-up.
b. Full mode: Claude, ChatGPT, Gemini, Qwen, Kimi and Grok each write their own answer, and so does my agent. They grade each other's work blind, never their own.
A model that didn't write the top answer turns it into the final version. Then Grok, whose job is to disagree, tears it apart before I see it.
Fast takes a few minutes and costs cents. Full takes about half an hour and a few dollars, so I save it for the big calls.
Everything's moving so fast so it'll probably end up looking different next year.
Anyways I spent a lot of time testing diff things out and reached a setup I'm very happy with.