Nous launches Hermes Desktop for self-improving AI agent

Nous Research released Hermes Desktop as a public preview on June 2, offering a native graphical app for macOS, Windows and Linux. The build is published as version v0.15.2 under the MIT license and is available for free download, inspection and modification.

Hermes Desktop provides a graphical interface for the Hermes autonomous agent and removes the need to run command-line install and configuration steps. The desktop app runs the same agent core as the command-line version and retains the same memory, skills and configuration so existing projects remain compatible.

Hermes is designed to learn from use. When the agent discovers an effective method for a task, it saves that method as a reusable skill document that it can apply automatically in later tasks.

The app is built on Electron with a React front end and a Python backend. Native builds are provided for macOS 12 or later and Windows 10 and 11 as direct downloads. Linux users currently install via the terminal.

Core features include persistent project memory, natural-language scheduling for recurring tasks and reports, web browsing, image generation, and access to more than 300 models through the Nous Portal.

The Nous Portal offers a free tier and paid plans named Plus, Super and Ultra. Paid plans include monthly credits and access to additional models.

Hermes Desktop supports delegated sub-agents, which are smaller agent instances that can run isolated tasks in parallel. The app also integrates with messaging services including Telegram, Discord, Slack, WhatsApp, Signal and email.

For task execution, Hermes can run in sandboxed environments using one of five backends: local, Docker, SSH, Singularity and Modal.

The project’s source code is available on GitHub under the MIT license, allowing users to audit, self-host or modify the code. The Nous Research team is collecting feedback while the public preview is live.

Prior to the official app, graphical interfaces for Hermes were community-built third-party projects. A competing open-source agent, OpenClaw, had previously shipped with a visual interface and a skill marketplace along with multiple messaging integrations.