Note: This is a re-post of my LinkedIn article.

I didn’t know it would cost this much!

I had a recent conversation with one IT Director who had an intern burn their entire monthly quota of ChatGPT credits in 2 days. An angry, C-suite executive called him complaining that ChatGPT wasn’t working anymore. He had to quickly upgrade the plan to keep things going.

These stories are common, but they don’t have to be. I keep hearing people and companies exclaiming about the costs of their AI models. Most harnesses like the Codex app from OpenAI, Hermes Agent, or Claude Desktop are free – the models are where almost all of the cost lies.

Avoid Vendor Lock-In

This is the same old tale of vendor lock-in that the IT industry has dealt with for years. You pick a vendor (VMware, Microsoft, OpenAI, whatever), get your discounts, and build your IT processes around them. It’s a mistake. It’s also a mistake to not consider open source, or in the world of LLMs, open-weight models. Things are moving fast, adaptability is key!

I’ve been using OpenClaw and Hermes Agent for many months now. It’s not to generate revenue and no one is sponsoring me, it’s for learning and hobbies. That means costs need to be very low. What I’m doing at home is what anyone and even any company can do as well. Let’s take a look at proprietary models and their open weight/value competitors:

Here’s the takeaway: After 30 days, operating costs dropped 87% and output quality only fell 4% on average – same revenue. That’s HUGE!

What About Data Theft

Here’s the #1 concern I hear, “I’m afraid of data theft or the Chinese companies training on my data.” Yes, real concern for companies but it’s easily addressable.

Many providers have configurable options for ZDR (Zero Data Retention) and to exclude your requests from being trained on. You can get most of these models under your existing company agreement with AWS or Azure, in fact. Both platforms have very strong guardrails that can be configured. Here’s a screenshot of my guardrail configuration for my health agent:

Some platforms that offer these models simply enable ZDR automatically. Here’s a snippet from OpenCode’s Privacy page:

The plan is designed primarily for international users, with models hosted in the US, EU, and Singapore for stable global access. Our providers follow a zero-retention policy and do not use your data for model training.

So, it’s not in China, not training on your data, and your data has zero-retention policies applied to it. Same kind of things are configurable on other platforms.

Balance It Out

You don’t have to pick one or the other. I have the ability to switch to Claude Sonnet or Opus at any time for certain tasks. It’s been a long time since I’ve had to but if my agent is stuck in a loop or just making too many mistakes, I can always pay-as-you-go on a Claude model.

The important thing is to use a tool that allows this. I use Hermes Agent as my daily driver; others have configured Claude Desktop to be able to do it (might break at any update) or use something like Goose or OpenWebUI – there are dozens of front ends.

Skip the manual switching entirely by using an LLM router/gateway. I’ve mentioned Manifest previously but recently Routerly caught my eye. Routerly has built-in smart routing and you can host it on-premises so your data stays yours.

Dismissing the open source (e.g., MiMo), open weight (e.g., Kimi, GLM) and low cost (Qwen is proprietary but it’s cheap) models is just costing you unnecessarily. Just because they don’t have the ad spend, US presence, and sales reps bugging you doesn’t make them less good. I’m almost exclusively using these models for coding, planning, writing, researching, and pretty much everything else we all use agents for today.

If you’ve tried the most recent versions of these models, what do you think? Are they on par with ChatGPT/Codex and Claude?

Note: This is a re-post of my LinkedIn article.

Overview

I’ve written about using OpenClaw, Hermes, and experimenting with other autonomous agent harnesses (like Claude Desktop). There is no “best architecture” because these things should be purpose-built for your way of working. In this article, I’m sharing how my setup has evolved using a few more tools and I’ll explain what they do.

The Harness

Hermes wins here because it’s self-improving, has great speed and writes its own skills. My agent needs to act like an employee or business partner, OpenClaw forgets too much. I still run an agent on OpenClaw because that’s how I started but going forward, new agents are built on Hermes Agent. An agent is only as good as the models it has access to. I’ve tried “free” tiers of models, and it makes the agent feel dumb. It will trick you because some things will work, you’ll get good responses, even a website built out for you, but you’ll end up regretting it – the quality is terrible.

The Memory

If I had a new install of Hermes, one of the first things I would do is give it better memory. It’s OK out of the box but there are plugins that make it better. I’ve heard honcho is a great choice but I’m running Mnemosyne. Mnemosyne is built for Hermes and allows you to dig into the memory system – not by analyzing a bunch of files or database tables, visually too:

With memory handled, it’s time to review models.

The Model

When using OpenClaw at first, I was switching between Claude Sonnet and Claude Opus, depending on my needs. These models are expensive, or if you’re using a subscription plan, time-consuming because you have to wait a while for quota resets. I’ve tried Z.AI subscription and then OpenAI Codex, too. I actually had the best experience with Z.AI‘s GLM-5-Turbo but sometimes I needed to switch. All of this switching is painful. I also tried various ways to incorporate a “model router” within OpenClaw and found them to be very clunky. I often got errors. So here’s my recommendation: use a multi-model subscription. There’s a few good subscription and pay-as-you-go options:

• OpenRouter – PAYG but every model is available.
• OpenCode – Subscription with 12 curated models.
• GitHub Copilot – Subscription (currently paused) with models from Anthropic, Google, OpenAI and more.
• Together AI – PAYG High-performance open-source models from Gemma, Mixtral, Qwen, and Llama.
• Fireworks AI – PAYG Fastest multi-model, similar models to Together AI
• Groq Cloud – PAYG Low latency Llama and Mixtral.

Best of all, you can use multiple of them but then switching between models and now model providers becomes really hard. We have a solution for this, that I explain in the next section. A quick note on privacy: some people distrust Chinese models, others distrust all cloud-hosted models. I get it and it may depend on the kind of work you’re doing. You can adapt these learnings to your posture.

The Routing

This was the key unlock for me most recently. We need a fool-proof way to route. Enter, Manifest. Manifest is an open-source LLM router for agents. It sends each query to the cheapest model that can handle it, saving up to 70% on inference costs. You can control the routing and set limits. But, best of all, it will automatically fallback to a different model on failure.

When I was using just OpenRouter, I often had model failures. OpenClaw and Hermes would try their best to re-try with their configured fallback model, but it was hit-or-miss (mostly miss). I was hand-editing config files to change the model and restart my work. Manifest fixes this. Here is my current routing config on Manifest:

Manifest can connect to your OpenCode subscription, OpenRouter, Anthropic, GitHub, Google, Minimal, Ollama Cloud, OpenAI, xAI, Z.AI, Deepseek, Moonshot, Nvidia and others (including custom). Then you can configure the routing, limits, which agents use what models, and other setting. Manifest also gives you a nice overview, here’s my token usage for the past few days:

I can see every query, which model was used, its cost, token usage, latency, and cache hit. In the next image, you can see one of the rows shows “Handled” meaning the model (Kimi K2.6, in this case) failed but Manifest sent it to Deepseek V4 Pro (fallback).

In Hermes, no errors, maybe a delay that I didn’t notice. Distrust cloud? You can run your own Manifest instance, using local models if you want.

The Architecture

Now that we have the pieces here’s the end-to-end architecture, to make this really simple.

1. Run Hermes in a VM, VPS, or on your own computer. Since I already had a computer lab, I just spun up a VM to host it. I like that all my files (memory, skills, etc.) are all saved locally.
2. The ONLY endpoint (cmd: hermes model) that Hermes needs is the one to Manifest and your API key. Manifest will handle the rest.
3. Within Manifest, setup any one of the subscription or pay-as-you-go plan or better yet, a combination of them. Setup routing, limits, and other settings. I’m using OpenCode Go.
4. Do yourself a favor and avoid trying to juggle free models. Pay for OpenCode Go, GitHub Copilot (when its available again) or add credits to OpenRouter to access frontier models (even if they are Chinese or open source).
5. For extra monitoring, use Langfuse. I have it setup but find that I rarely need to go there anymore since I am using Manifest.

If I started heavier work and ran into subscription limits within OpenCode, my next step is either a subscription for MiMo, Kimi or GLM – all are low cost and really good. I mean, just look at this 99% price reduction from Xiaomi:

The Wrap

Claude Opus and GPT-5.5 get all the hype but I see people struggling with Hermes (or OpenClaw) because they picked one of these models and think they’re set. Here’s an example:

Is Hermes the problem? Is GPT-5.5 the problem? There will be problems, build a resilient architecture!

Note: This is a re-post of an article I published on my LinkedIn.

Innovation doesn’t wait for you

Every day, my feed filled with new OpenClaw clone, new frameworks, so many “this changes everything” posts. But, I was already OpenClaw on a VM and it was working fine. It was powerful and I made my customizations. So, I ignored the noise and resisted change.

Then Nous Research dropped Hermes Agent v0.6 with multi-agent support. That was the nudge I needed. I had previously tried Paperclip.ai for multi-agent orchestration and it just didn’t work for me.

So many claws

OpenClaw started this wave but the ecosystem is exploding. Some notable clones:

• IronClaw → Rust-based, more enterprise-ready
• ZeroClaw → minimal, faster, simpler
• NemoClaw → Nvidia’s security-focused version
• And many more: PicoClaw, MetaClaw, CoPaw, NanoClaw, NullClaw…

I’m sure there are many more. Development on OpenClaw moves fast, new major features are being released on a weekly basis. That’s something that typically takes months. We’re talking about things like agent capabilities, toolsets, critical security patches and workflow features like /tasks that change how you work. The question remains: do any of these new “claws” substantially improve things for my main issue? The main issue being memory.

The real problem: memory

I was running 2 agents:

• Agent 1 helps me run my home systems (Proxmox, HomeAssistant, Security cameras), social media, and research.

• Agent 2 is more personal, calendars, emails, reminders, birthday tracking and gift ideas.

It worked… until it didn’t. My main agent started getting confused. Like a “jack of all trades” employee with too many responsibilities and not enough specialization. To fix it, I added:

• lossless-claw → store everything
• ByteRover → find everything

It helped. But it still wasn’t great. Why? Because I was doing the work the agent should be doing. Things like telling it to remember, verifying that it could retrieve the facts, and managing plugins. At some point, you realize that the more plugins you add the more you become the orchestrator.

Hermes fixes memory at the system level

Here’s a comparison:

• OpenClaw: memory is external (bolt on), plugin-dependent, and fragmented.
• Hermes: memory is built-in, it’s selective and contextual and most importantly self-maintaining.

Let’s break it down:

Feature	OpenClaw	Hermes Agent
Persistence (storing facts)	needs lossless-claw or similar plugin	built-in
Retrieval (finding context)	needs ByteRover or other plugin	built-in
Synthesis (making memory usable)	not really there yet	core feature
Autonomous updates (deciding what matter)	manual	automatic

OpenClaw tries to remember everything. Hermes tries to remember what matters. That means less babysitting, better decisions and much cleaner execution.

Skills are the real unlock

Better memory was great by itself but then I noticed something wild! Autonomous skills.

Hermes started creating skills… on its own. No prompt, no instruction. For example, I asked it to fix a broken local site. It SSH’d in, fixed it, then created a reusable skill. I didn’t ask it to do that. Now it knows where the files are (/opt), the stack (NextJS, Tailwind) and how to fix it again in the future.

That’s not just automation. That’s self-improvement.

As of Hermes 0.6, it can spin up subagents on its own, run tasks in parallel, and share memory and skills. This is a shift from tools to teams. Separate agents means better outcomes, less confusion, just like real specialized employees.

What held me back

I hesitated because OpenClaw has a huge ecosystem, tons of plugins, and a massive community. But it didn’t matter because both Hermes and OpenClaw can use Clawhub (skills directory) and support shared skill ecosystems. So the gap isn’t as big as it looks.

We already see enterprises adopting this paradigm. Microsoft moved here. Omar Shahine recently talked about bringing OpenClaw-style agents into M365. When Microsoft moves, it’s not a trend, it’s a signal.

If you’re technical, spin it up with Docker in minutes. If you’re not, use a hosted option. OpenClaw has many. Hermes? Still early. I only found one viable hosted route (Hostinger). Later, Nous Portal announced their own hosted options (this is where I would go).

Final thoughts

We’re not just watching better AI tools emerge. We’re watching a shift from assistants that respond to agents that do, remember, learn, and delegate. We’re getting closer to Iron Man’s Jarvis. OpenClaw showed what was possible. Hermes is starting to show what’s sustainable. The real question now isn’t which one wins. It’s how long before we stop doing this work ourselves.

This is a re-post, with edits, from my original post on LinkedIn.

Democratizing AI

I’ve been using OpenClaw for several months now, since it first was released as Clawdbot. And, honestly, this is what AI should have been all along. The sad part is that it took a developer tinkering as a hobby, then coming out of retirement, to create this project. A single developer. As someone whose career was spent at very large companies, including software and AI companies, it’s embarrassing!

ChatGPT was a marvel when it was released, you could chat with the entire knowledge of the internet. But now that feels like a small step up from Google search. I think AI was rushed and is still being rushed; to capture as much market share and profit – that’s how these things work unfortunately. For most people, it means their experience will be subpar and they’ll give up on it. It’s true, AI is more hype than real right now – especially in the walls of a company.

It reminds me that innovation often happens at the margins. It happens in silos or small communities or a humble startup. It happens increasingly in open source. Large companies like Microsoft (where I spent more than a decade), Amazon, Google and others simply don’t give space for innovation to happen. Interestingly, these companies were once where innovation did accelerate – but they got too big, too bureaucratic, too risk averse.

It’s for that reason that I think that Enterprise (large companies) will not drive AI innovation – it moves too fast. Microsoft was talking about “agentic experiences” over and over, ad nauseum, while Peter Steinberger built and shipped it, by himself. OpenClaw is now the most starred project on GitHub, higher than Linux, python, React and well, anything from a large software company. Take a look at the red line – that’s not the border, it’s the OpenClaw repository:

What is OpenClaw?

You might be wondering now, what in the world is OpenClaw, especially if you live under a rock (like most of my LinkedIn connections who operate in the “Enterprise” space 😊). OpenClaw democratizes agentic experiences, it brings the party to you. It’s not a separate app, it’s the chat app you already use: Telegram, Discord, iMessage, and dozens more. And it’s not a single agent; it can be, but it can also be a swarm of agents. And it’s not just reactive (you ask, it answers), it’s proactive!

I’m simple and slow, so let me explain a few things I’m doing. I want it to be my personal assistant. I gave it a name, a personality, a soul and the ability to learn about me. It tells me what’s on my calendar, it can manage it. It’s scouring the web for deals for RAM (I need to upgrade) and sends me a link to eBay when a deal pops up. It’s now managing my Proxmox virtual host, all of the virtual machines and all of the containers. I get a daily report on CPU, RAM, disk space, if backups fail, if something goes wrong and if it was able to fix it for me while I was asleep. It’s posting on social media for me. When I ask it to “vacuum the house,” it starts my Roborock vacuum cleaner – even if I’m miles away from home. It has all the knowledge from my Notion database. It built me a website, found a free tier and pushed code to it. It takes care of itself, it cleans up, it backs up itself and it commits things to memory. I don’t write code – I just ask it, like I would ask an employee, and it goes to work.

But as I said, I’m pretty simple. There are thousands of skills you can add. Skills to enable voice (like Alexa), enable calling real phones and speaking to the person on the other end, enable using credit cards and crypto to buy things for you. Skills to automate content creation, video creation – indeed people are running their entire businesses with it. People are making their “claw” trade crypto, futures, or prediction markets. And non-developers are building polished, amazing apps. People are using it send flyers in the mail or cold call leads, all automatically. There’s now an entire ecosystem around OpenClaw and competing solutions, like Hermes Agent and Nvidia’s NemoClaw – and you don’t need to be technical at all to get started.

I think the best way to get started is to pony up some money (yeah, there’s free ways to do it but more complicated) and get a hosted OpenClaw instance. I like MaxClaw — Cloud-Hosted AI Agent by MiniMax ($40/mo.) or Clawi.ai – Your Personal AI Assistant in the Cloud ($30/mo.). Just don’t go out and buy a Mac Mini for a few hundred bucks and try to do it yourself the first time – trust me. You can migrate to that later if you want to and PLEASE start with using a frontier model like OpenAI Codex or Anthropic Claude Opus/Sonnet – it makes all the difference in the world.

My message to you if you live in the corporate world: step outside it for just a bit. The consumer world will win AI. The Enterprise will have their friendly, guard railed, watered-down version to create ugly PowerPoints or analyze spreadsheets worse than an intern. But you could be using it now for your personal life or your business, side gig, or just for fun.

Are you using OpenClaw or something like it? Why or why not?