Grok Bot WAS Brilliant. That's Why I'm Moving to Hermes.

Nov 7, 2026 · 10 min read · original on X

A few days ago, I published an article explaining how I keep my Grok Bot usage low without sacrificing capability. Before that, I wrote about connecting Grok Bot to Hermes so my two AI agents could work together.

I still stand by both articles.

But after watching what’s happened over the past week, I’ve reached a different conclusion.

If you’re serious about using AI agents for your business or everyday life, you shouldn’t let one company control the entire system you’re building.

And that’s why I think it’s time to start moving away from Grok Bot for anything important and toward something like Hermes Agent.

Grok Bot really was that good 🤖

I want to give credit where it’s due.

When Grok Bot launched, it was genuinely impressive. Not just another chatbot with a fancy interface, but something that finally made the promise of personal AI agents feel real.

You could create specialized bots, give them responsibilities, connect your accounts, schedule routines, and have them work on real tasks. They had persistent computers, could use applications, and could coordinate work without you having to explain everything from scratch.

The interface was polished. Getting started was easy. And the intelligence behind it was, frankly, pretty remarkable.

I built my own small team with a coordinator, an IT specialist, and other purpose-built assistants. I could interact with them from my phone or computer. It was the kind of setup that would have required a ridiculous amount of tinkering not long ago.

That’s what made Grok Bot exciting.

It made something previously reserved for technical enthusiasts accessible to ordinary people.

And I think SpaceXSI deserve credit for that.

But great software and a great product relationship aren’t necessarily the same thing.

Then somebody moved the goalposts 😡

Over the Halloween weekend, Grok Bot users started reporting something strange.

Their weekly usage meters were suddenly jumping. Some woke up to find their allowances completely exhausted, even though they hadn’t been doing anything unusual.

Some of those initial readings were subsequently corrected. But other users started digging into their actual usage records and found something potentially more concerning.

They appeared to be getting substantially less work out of the same allowance.

Posts started showing up with similar complaints.

These are individual reports, not independently audited company-wide figures. But they aren’t simply people complaining that a percentage looks different.

They’re comparing actual work and actual usage.

On November 1, Elon Musk acknowledged that serving Grok Bot had become unsustainably expensive relative to revenue.

I understand that. AI inference isn’t free. Agents can consume enormous amounts of compute, especially when they’re running tools and working autonomously.

I’ve never expected unlimited access for a few dollars forever.

What I do expect is to know what I’m paying for, and to have a reasonable expectation that the service won’t substantially change overnight without notice.

That’s where this starts feeling like a rug pull.

The subscription price hasn’t necessarily changed, but for some users, the amount of useful work they’re getting from it apparently has.

And when your business depends on that work, that’s a much bigger problem than an annoying usage meter.

This is already affecting real work

It’s easy to dismiss complaints about AI usage as people wanting more tokens to play with.

But take a look at examples of what actual users are reporting.

These aren’t theoretical concerns about what might happen someday.

These are people who organized real work around a service, only to find that the terms they were relying on had apparently changed.

Think about what that means for a solo business owner.

You might have spent weeks teaching your bots how to handle customers, maintain your website, prepare reports, research opportunities, or manage repetitive administrative work.

Then one morning, you discover that the workflow you built around yesterday’s service no longer fits today’s limits.

You didn’t change anything. Your work didn’t change.

The platform did.

And now people are saying Grok Bot got dumber 🥴

This is the part that concerns me even more than the usage limits.

Across multiple discussions, users are reporting that their bots suddenly seem less capable.

Tasks that previously worked on the first attempt now require corrections. Bots appear to forget things they previously remembered. Some are making mistakes in workflows that had been running reliably.

Many folks across 𝕏 report their bot making up details, cutting corners, and struggling with work it had previously handled.

Did SpaceXSI secretly replace the underlying models with cheaper or less capable alternatives?

I don’t know. And nobody outside the company has publicly demonstrated that’s what happened.

There are other possible explanations, including changes in model routing, context handling, tool behavior, or ordinary service degradation.

But here’s what makes the situation interesting.

Grok Bot’s own documentation explains that it can route requests to different models, including third-party models. It also explicitly says that users cannot see which model handled a particular request or choose the model themselves.

Read that again.

You can build your business around an AI agent without actually knowing which model is doing the thinking.

And you can’t lock it to the one that was giving you good results.

Even if the recent intelligence complaints turn out to have another explanation, that lack of transparency remains.

I don’t want the intelligence my business depends on to be a mystery setting someone else controls.

This is why Hermes Agent makes so much sense

I’ve been running @NousResearch Hermes Agent on my own lab server, and it’s becoming increasingly clear why this is the direction I want to go.

Hermes isn’t an AI model. It’s the software that runs and coordinates your AI agents.

Think of the model as the brain and Hermes as the system that gives that brain its tools, memory, instructions, and ability to actually do things.

It has persistent memory, skills, scheduled tasks, browser automation, terminal access, integrations, and the ability to delegate work to other agents.

It can learn reusable procedures from experience rather than having to rediscover the same steps every time.

And it isn’t limited to one company’s model.

You can use OpenAI, Anthropic, xAI, OpenRouter, Nous, and other supported providers. You can also connect your own locally hosted models.

That’s the important difference.

I get to choose which model does the work, rather than having a company make that decision behind the scenes.

If a provider raises prices, changes limits, or becomes less attractive, I can change providers without rebuilding my entire agent system.

If a particular model performs better at coding while another handles research more effectively, I can make those choices myself.

And if I eventually have the hardware to run capable open-weight models locally, I can reduce my dependence on paid inference even further.

To be clear, local models aren’t automatically as intelligent as the best frontier models. And using Hermes with a cloud provider doesn’t magically eliminate that provider’s prices or limits.

But it changes something fundamental.

The agent system is mine. The model is replaceable.

That’s a level of independence Grok Bot doesn’t currently offer.

Hermes has also come a long way

I’ll acknowledge something else.

Grok Bot had a huge advantage in presentation and ease of use. Hermes historically required more technical knowledge, and setting up your own infrastructure isn’t everyone’s idea of a fun Saturday.

But that’s changing.

Hermes v0.23.0, released November 5, includes a much more approachable desktop experience, Bot Mode with individual agents and their routines, group chats, and a web interface you can serve from your own gateway.

It can manage multiple machines and use plugins to expand what your agents can do.

It’s still not as effortless as signing into a polished, fully managed cloud service. You have to think about hosting, backups, access controls, updates, and which models you’re paying to use.

But those tradeoffs are becoming much easier to justify.

Especially when the alternative is building increasingly important workflows on a platform whose capabilities and limits you can’t independently control.

The creator of Hermes just made leaving Grok Bot easier

Here’s something particularly timely.

On November 2, @Teknium the creator of Hermes Agent, merged pull request #147724, adding a migration path specifically for Grok Bot users.

Hermes can now help import your existing bots’ roles, memories, skills, and scheduled routines rather than requiring you to recreate everything manually.

That’s a pretty big deal.

Grok Bot doesn’t currently provide a normal export button for all that information, so Hermes uses an interesting workaround: it generates a prompt that asks your Grok Bot to export its own configuration.

Here’s the basic migration process:

1. Install or update Hermes Agent. Run hermes setup for a new installation, or hermes migrate grok-bot if Hermes is already configured.
2. Export your Grok Bot setup. Hermes provides a prompt you paste into your Grok Bot’s chat. The bot generates an export file containing the supported information about its setup.
3. Import it into Hermes. Download that file and use hermes import-agent grok-bot --source grok-bot-export.json, reviewing the proposed changes before accepting them.
4. Reconnect and verify everything. Scheduled routines are imported paused to avoid duplicate execution. Event-triggered routines need to be rebuilt, and connected accounts need to be authorized again. Test your workflows before retiring the originals.

This isn’t a magical one-click clone of your entire Grok Bot account. Some things need manual attention, and you should review what your bot exported.

But compare that with having to rebuild every role, instruction, skill, and routine from scratch.

It’s a much better starting point.

And importantly, migrating doesn’t require immediately abandoning the service you’re already using.

You can move a few workflows, test them, and gradually reduce your dependence.

That’s how I’m approaching it.

Owning your agents doesn’t mean never paying for AI again 💳

I think this point gets lost whenever people talk about self-hosting and open source.

The goal isn’t necessarily to eliminate every subscription or API bill tomorrow.

The goal is to stop being trapped by them.

Maybe I use a powerful frontier model for difficult research and decision-making. Maybe I use a cheaper model for routine administrative work. Maybe a local model eventually handles most of the repetitive tasks running on my server.

I’d rather pay for the intelligence I actually need than have one company decide how much intelligence I get this week.

There’s also an important responsibility that comes with running your own system. You need sensible permissions, secure handling of credentials, backups, and safeguards against agents making costly mistakes.

Freedom doesn’t eliminate maintenance.

But I’d rather have that responsibility than discover that months of work depend on business decisions I have absolutely no influence over.

Don’t wait for the next rug pull 🫥

This isn’t really about Elon Musk, SpaceXSI, or even Grok Bot.

We’ve seen subscription services change pricing, limit features, and quietly adjust what customers receive for years. AI companies aren’t immune to those pressures.

In fact, the economics of running powerful AI models make those pressures especially obvious.

I don’t think every company making an unpopular pricing decision is acting maliciously. Sometimes the business genuinely can’t sustain what it originally offered.

But understanding why something happens doesn’t make the consequences disappear.

And the lesson for consumers should be pretty simple.

Don’t build something important around a service you can’t replace.

For me, that means keeping my agents’ instructions, memories, skills, automations, and work product somewhere I control, while retaining the freedom to choose the models powering them.

Hermes isn’t perfect, and it won’t solve every problem. But its open-source approach gives me a foundation I can maintain and adapt, instead of forcing me to rebuild my working life whenever a provider changes direction.

Grok Bot proved just how valuable personal AI agents can be.

I still think it deserves recognition for that.

But the past week has also shown what happens when that value is completely dependent on someone else’s platform.

I don’t want to rent the entire worker anymore. I’d rather own the system and pay for the brain when I need it.

That’s why I’m moving my workflow toward Hermes.

And if you’re building your business or daily routines around AI agents, I think you should seriously consider doing the same.

Because here’s the question worth asking:

If your AI provider cut your usage in half tomorrow, could your work keep moving—or would you be starting over?

Originally posted on X.