It learns as it goes
After a complex task, it writes a skill describing how it did the work, then refines that skill with use. It also keeps a curated memory and can search past sessions.
Compare
Hermes Agent
an open-source AI agent from Nous Research that runs on your machine or a server, remembers your work, and turns completed tasks into reusable skills
Who they are and what they are good at
Hermes Agent is an MIT-licensed agent from Nous Research, the lab behind the Hermes family of open models. Nous calls it a self-improving agent, and that is the core idea.
It runs where you choose
A terminal app on Linux, macOS, WSL2, and Windows. Commands can run locally, in Docker, over SSH, or in cloud sandboxes such as Modal and Daytona.
It reaches you where you are
One gateway serves Telegram, Discord, Slack, WhatsApp, Signal, email, and the command line.
Tools and models
Dozens of built-in tools, including a terminal and a browser, plus MCP servers, a built-in scheduler, and subagents for parallel work. Models come from Nous Portal, OpenRouter, OpenAI, or a custom endpoint.
Hermes is a good fit for developers and researchers who want an agent that gets better at their kind of work over time, on hardware they control.
Where Hermes stops for messy back-office work
Hermes is designed to change itself. That is a strength for open-ended work and the opposite of what a monthly filing needs.
The procedure can rewrite itself
Skills it writes are refined during use. For a tax return, the steps should change only when someone approves the change.
Safety is a setting
The docs describe command approval with smart, manual, and off modes, a session-wide bypass, a hard blocklist, and default-deny sender pairing. The local backend runs commands on the host, and the docs recommend a container backend for production. They also say its write denylist "does not sandbox a hostile or compromised agent."
Built for the terminal
Setup, configuration, and skills assume someone comfortable with a command line. A desktop app with no API is not its focus.
Exact math is model judgment
Tying a payroll register to the bank debit to the cent should not depend on a model.
Evidence is a transcript
Session history shows what the agent said and ran, not the screen it changed.
LaunchAI builds each procedure as fixed steps that are checked live while you build, runs them on your machine with your existing logins, and keeps a screenshot of every action. See Product →
Side by side
| Hermes Agent | LaunchAI | |
|---|---|---|
| What it is | An open-source, self-improving agent | A desktop product for recurring procedures |
| Who builds | Developers and technical users | The process owner, by talking and showing |
| How it changes | Writes and refines its own skills | Only when a builder edits and re-validates a step |
| Reaches software through | Terminal, browser, MCP servers, and skills | The screen, so any app on the machine, with no integrations |
| Where it runs | Your machine, a server, or a cloud sandbox | Your machine or an enrolled agent PC |
| Exact math | Model judgment | Deterministic rules, no model call |
| Human control | Command approval modes and sender pairing | A review step anywhere in the workflow |
| Scheduling | Built-in scheduler that reports to chat | Schedules, a run console, and review queues |
| Evidence | Session history | Before-and-after screenshots of every action |
| Cost | Free software, plus your model account or a Nous Portal subscription | License, plus your model account |
| Support | Community | Human support |
When to choose Hermes
- A researcher wants a long-running assistant on a server that remembers past projects and gets faster at recurring research chores.
- A developer wants an open agent to extend with MCP servers and their own skills.
- A lab wants to generate tool-use data for training its own models, which Hermes supports directly.
When to choose LaunchAI
- Every pay period, pull the payroll register, tie it to the bank debit to the cent, and hold any difference for the payroll manager.
- The steps must stay the same until someone decides to change them.
- The apps live on a Windows desktop with no API, behind logins you already have.
- An auditor will ask what happened, and the answer should be a screenshot.
Better together
Hermes researches, LaunchAI enters. Hermes drafts findings into a shared spreadsheet. A LaunchAI agent picks up each approved row, re-checks it on the live screen, enters it, and routes anything unusual to a person.
FAQ
Is Hermes Agent free?
Yes. It is MIT-licensed open source. Nous Portal is an optional paid subscription for models and some hosted tools, or you can bring another provider.
Does Hermes Agent only work with Hermes models?
No. It works with many providers, including OpenRouter, OpenAI, and custom endpoints.
Does Hermes ask before running risky commands?
It can. Approval modes include manual and smart, but they can also be switched off, and checks are skipped inside container backends, where the container is the boundary.
Does LaunchAI learn on its own like Hermes?
No, by design. A LaunchAI workflow changes when a builder edits a step and validates it live. That is what makes a run repeatable.
No prompts. No babysitting. No Integrations.