Why Hermes Isn’t Replacing Claude Code (Yet)
A 205,000-star GitHub repo, free frontier models, and documentation that will make you question your life choices.
Hermes is simultaneously the most frustrating and the most interesting AI tool I’ve used. Just when you’re ready to uninstall it, it does something that reminds you why 205,000 people starred it on GitHub.
If you’re wondering whether those stars are deserved, this article should save you a week of trial and error.
What Hermes Is (The Short Version)
Hermes is less like a chat-bot and more like an AI operating system. It remembers previous work, builds reusable skills, delegates tasks to sub-agents, and lets you continue the same conversation from your terminal, web browser, or messaging apps.
It’s developed by Nous Research, but despite the name, Hermes Agent isn’t tied to the Hermes language models. It works with any compatible model, whether that’s Claude, DeepSeek, Qwen, or something else.
What Makes Hermes Different
Hermes isn’t trying to be a better chat-bot. It’s trying to be an AI workspace where agents work together.
Instead of throwing everything into one giant prompt, it breaks complex problems into smaller tasks, delegates them to sub-agents, and stitches the results back together. That keeps context focused instead of endlessly growing.
The little things are what sold me. I can switch from Claude to Qwen halfway through a conversation with /model without losing context. I can start a task in my terminal and pick it up later from Telegram or Discord. And every now and then, Hermes quietly reuses something it learned days earlier, saving me from explaining the same thing all over again.
It’s not magic. Sometimes you still have to remind it to use those skills, and the rough edges are very real. But when everything clicks, it stops feeling like a chat-bot and starts feeling like you’re managing a small team of slightly chaotic interns.
My Setup
I use NyArch Linux (btw). OpenRouter is my primary provider, with NVIDIA NIM for models I can’t get elsewhere. No local GPU for inference — summer is warm enough here already.
The appeal is flexibility. OpenRouter’s free tier includes surprisingly capable models like Owl-Alpha, while NIM gives free access to DeepSeek V4 Pro, Qwen 3.5 397B, Kimi 2.6, MiniMax 2.7, and more. DeepSeek V4 Pro in particular punches well above its price tag of “free,” as long as you can live within the rate limits.
My goal was simple: build a provider-agnostic AI workspace that could route each task to the model best suited for it.
That’s one area where Hermes genuinely shines.
Installation
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bashThe installation was refreshingly uneventful. No dependency hell, no mysterious Python errors, and hermes doctor came back clean.
For the first twenty minutes, Hermes felt almost magical. The CLI worked, memory worked, tools worked, and I genuinely thought, “This is going to be easy.”
One piece of advice: skip Hermes Desktop. It’s still rough around the edges, feature-incomplete, and ironically has to be launched from the terminal anyway. The web interface is far more polished — it’s where the Kanban board really shines — and Hermes One is a good alternative if you prefer a more traditional chat interface.
Then you’ll try to do something the documentation didn’t expect that’s when the honeymoon end.
The Documentation Problem
The documentation is excellent — right up until you leave the happy path.
Hermes has a surprisingly deep feature set: multiple providers, fallback chains, sub-agents, MCP servers, gateway integrations, and an OpenAI-compatible endpoint. The problem isn’t that the documentation is missing. It’s that it often tells you what exists, but not how to make it work.
NVIDIA NIM was the clearest example.
It’s listed as a supported provider, but getting sub-agent delegation working required manually adding a provider block to each profile’s config.yaml. The hermes model wizard doesn't do it, and the resulting 400: model not found errors don't point you anywhere useful.
I spent almost three hours convinced NIM was broken before discovering the problem was a missing configuration block.
Here’s what finally worked:
providers: {}
fallback_providers:
- provider: custom
model: qwen/qwen3.5-397b-a17b
base_url: https://integrate.api.nvidia.com/v1Once I had the fix, everything worked exactly as expected.
That was the frustrating part. The software wasn’t broken — I just couldn’t find the missing piece in the documentation.
Provider Parity Is the Real Issue
The missing provider block was frustrating, but it also exposed a bigger problem.
Hermes promises model flexibility, yet the experience clearly varies between providers. Claude users are likely to have a much smoother on-boarding than anyone relying on custom endpoints or less common providers.
Get SamaCorpInc’s stories in your inbox
Join Medium for free to get updates from this writer.
If a provider is officially supported, core features like sub-agent delegation should work without manual configuration. That’s not a documentation issue — it’s a product issue.
What Actually Impressed Me:
The Kanban Board
I expected the Kanban board to be a gimmick. It ended up being one of Hermes’ best features.
Watching the agent break a complex task into smaller sub-tasks, assign them, and track their progress makes long-running work feel far more manageable. You’re not staring at a spinner wondering what the model is doing — you can actually follow its reasoning as the work unfolds.
A couple of lessons I learned the hard way.
Skip the built-in orchestrator and ask the agent to build the task tree itself in a normal chat. You get more control, and the results are usually better.
Also, set a concurrency limit. Hermes will happily launch as many sub-tasks as you let it. I learned that by accidentally exhausting every free-tier rate limit I had within seconds.
Context Management
This is where Hermes quietly beats Cline for long-running development work.
Instead of constantly growing the context window, Hermes actively manages it. Older conversations get compressed, side tasks like vision or web summarization are handed off to cheaper models, and only the information that still matters stays in front of the primary model.
The result is lower token usage without a noticeable drop in quality. The system is managing context on purpose instead of simply hoping a larger context window will solve the problem.
Third-Party Compatibility
Hermes doesn’t try to lock you into its own ecosystem. Between its built-in tools and MCP support, it can integrate with just about anything you throw at it.
Not every MCP server worked out of the box, and I still ran into configurations the documentation didn’t cover. But once connected, the flexibility is impressive. It feels like a platform you can grow into rather than one you’ll eventually outgrow.
It Actually Remembers
Most AI tools say they have memory. Hermes was the first one that made me notice it.
The Rough Edges
Hermes isn’t polished yet, and you’ll notice that quickly.
The architecture is impressive. The on-boarding isn’t. Once you leave the happy path, the documentation gets thinner, some integrations need more manual configuration than they should, and you’ll occasionally end up reading source code to understand why something isn’t working.
The frustrating part isn’t that these problems exist. It’s that they’re standing in the way of software that’s otherwise remarkably capable.
Who It’s For
If you’re looking for the best coding assistant, I’d still point you toward Claude Code, Cline, or Copilot. Their IDE integration is tighter, the feedback loop is faster, and they’re simply more polished for day-to-day development.
Hermes shines somewhere else.
If you’re reasonably comfortable with a terminal and want an AI workspace that follows you across devices, remembers previous work, switches models on the fly, and handles long-running tasks, it’s genuinely impressive. That’s where the architecture starts making sense.
I wouldn’t recommend it to complete beginners. Not because it’s difficult to install — it isn’t — but because you’ll eventually leave the happy path, and when you do, you’ll need to be comfortable reading logs, editing configuration files, and figuring things out as you go.
What I’d Like to See
Hermes doesn’t need more features. It needs polish.
The documentation should reflect how people actually use the project, not just the happy path. Custom providers, subagent delegation, and configurations like NVIDIA NIM should either work out of the box or be documented well enough that users don’t have to reverse-engineer them.
The OpenAI-compatible endpoint also deserves another look. Today, requests flow through the sub-agent system, so reasoning content gets stripped before responses are returned. That trade-off makes sense for agent orchestration, but it’s a shame for an endpoint meant to behave like a standard OpenAI API. I’d love to see those use cases separated in the future.
Finally, provider parity should become a first-class goal. One of Hermes’ biggest strengths is the promise that you can bring your own model. That promise only works if every officially supported provider receives the same level of polish — not just the most popular ones.
Verdict
Hermes didn’t replace Claude Code or Cline.
It replaced something I didn’t realise I was missing: a persistent AI workspace.
I still reach for Claude Code when I need to write code quickly. But when I’m researching, planning, coordinating long-running work, or simply moving between devices, Hermes has become the tool I leave running in the background.
It took less than an hour to install, three days to get working the way I wanted, and another two to understand why it’s built the way it is.
Would I do it again?
Absolutely.
Running Hermes on NyArch Linux with OpenRouter (owl-alpha mostly) and NVIDIA NIM (Qwen 3.5 397B, Kimi 2.6, MiniMax 2.7, DeepSeek V4 Pro). All configuration done manually — no Nous Portal subscription.
