Most "AI agent" projects are chat wrappers with a prompt template glued on. Hermes One is a different animal. It is a free, open-source desktop agent from Nous Research that writes its own skills after finishing complex tasks and keeps memory across sessions. It also runs scheduled jobs on Windows, macOS, and Linux. This review is written after several weeks of running it as a working tool, not a demo.
Short version: if you are a developer who wants an agent that compounds knowledge over time, this is one of the strongest open-source options in 2026. If you want something that installs and works in five minutes with no config, keep walking.
Key Features
The feature list on the site is long. Here is what actually matters in daily use.
Closed learning loop
After finishing a task with more than a few steps, the agent proposes a new skill file, writes it, and files it into its own skill library. The next time a similar task comes up, it picks up the skill instead of reasoning from scratch. This is the single reason to prefer Hermes One over a stateless chat wrapper.
Messaging gateway
One process, six inbound channels: Telegram, Discord, Slack, WhatsApp, Signal, and CLI. You point each channel at the same running agent and it treats every conversation as a session. This is genuinely useful if you want to hand a task off from your phone while away from the machine.
Cron scheduler
Built-in scheduler for unattended jobs. Daily reports, weekly backups, whatever you register. It uses the platform-native scheduler under the hood (Task Scheduler on Windows, launchd on macOS, systemd or cron on Linux), so a rebooted machine picks jobs back up.
Terminal backends
Seven execution targets: local shell, Docker, SSH into a remote box, Modal, Vercel Sandbox, and two more. Same skill, different sandbox. You can develop a skill locally, then run it against a Modal container for isolation when a client is watching.
Model-agnostic connection
OpenRouter (300+ models), OpenAI, Anthropic, Gemini, Ollama, and any custom OpenAI-compatible endpoint. One command switches providers. If Anthropic prices go up on Tuesday, you route the agent through OpenRouter's Claude on Wednesday.
FTS5 session search with LLM summaries
Every session is indexed in SQLite FTS5 and periodically summarized by the model itself. Ask "what did I do about the Postgres migration two weeks ago" and it returns the session with a short summary. It works. It is not magic. Recall drops on very old sessions if you never opened them again.
Subagent delegation
Parallel workstreams via Python RPC. You script an outer agent that spins up three sub-agents on independent tasks and joins their results. This is the closest thing in open source to Claude's Agent SDK pattern, and it runs against any model.
Pricing Breakdown
| Plan | Price | What you get |
|---|---|---|
| Free (only plan) | $0 | Full feature access, all 14 toolsets, unlimited model connections, self-hosted |
There is no paid tier. Hermes One is fully open source and self-hosted. Your real costs are model API bills (whatever provider you point it at) and whatever you spend on infrastructure to keep it running. On a laptop with a local Ollama model, cash cost is zero.
Pros and Cons
Pros
- Free and open source, backed by 14,000+ GitHub stars with active commits from Nous Research.
- The skill-creation loop is a real differentiator. Most competing agents forget everything the moment a session ends.
- Model-agnostic design means no vendor lock-in. Switch from Anthropic to OpenRouter in one command.
- Multi-platform messaging from a single gateway covers channels most agents ignore. WhatsApp and Signal are the interesting ones.
Cons
- Technical setup barrier is real. The origins are CLI, and it shows. Non-developers will bounce off the config on day one.
- The self-improving skill system is opaque. It is hard to audit what the agent has learned, and harder to make it unlearn a bad habit without wiping the skill directory.
- No managed hosting. You supply your own model API keys and your own machine.
- The desktop GUI wrapper is community-maintained. The Nous Research core and the GUI can drift out of sync between releases.
Who Is It For
Three concrete profiles that get value from Hermes One today:
- A solo developer who wants an agent that gets smarter about their own repos and workflows over months, not just within one chat.
- An operator running scheduled jobs across services, who wants one process to handle both cron and inbound messages from Telegram or Slack.
- A researcher comparing model outputs on the same tasks, who wants to swap Claude, GPT, Gemini, and a local Ollama model without rewriting orchestration code.
Three profiles it is not for:
- Anyone who wants a polished consumer chatbot with a sign-up screen. Look at Jan or Cherry Studio.
- Teams that need SOC 2, SSO, or a managed control plane. This is a self-hosted single-user tool.
- Anyone who wants to trust an agent's compounding skills without being able to read the code that generated them.
How It Compares
Quick landscape read. Jan and Open WebUI are stronger as local chat UIs with clean model management, but neither has the skill-persistence loop. AnythingLLM is closer on document-grounded workflows but weaker on agent autonomy and scheduling. AutoGPT was the original self-directed-agent idea but has lost momentum against tools with tighter learning loops. Cherry Studio is the best-looking desktop client of the group, and the worst fit if you want an actual agent with a scheduler.
Hermes One wins on two axes the others do not compete on: cross-session skill learning and unified messaging.
Verdict
Rating: 7.4 / 10.
Recommendation: install it if you are a developer and skip it if you are not.
Hermes One is one of the most serious open-source attempts at a persistent, self-improving desktop agent. The skill loop and the cross-platform messaging gateway do things no chat wrapper does, and the price is zero. The catch is honest: you pay for it in configuration time and model API bills. You also spend ongoing work reading what the agent has taught itself so you can prune the parts that go wrong. For a developer who wants an agent that gets more useful across months instead of resetting every session, that trade is worth taking. For anyone who wants an out-of-the-box tool, this is not the one.
