Nous Research has launched Hermes Desktop in public preview. This native app works on macOS, Windows, and Linux, giving the open-source Hermes Agent a proper visual interface. Previously, users could only interact with Hermes through a command-line interface or messaging platforms. The current version running behind the scenes is Hermes Agent v0.15.2.
According to Nous Research’s official documentation, the desktop version taps into the exact same agent engine. Your settings, API keys, conversation history, learned skills, and stored memory all sync between the desktop app and the CLI or messaging gateway. Think of the desktop as just another window into the same agent rather than a separate copy.
What is Hermes Desktop
Hermes Agent is a standalone AI agent, meaning it’s not a coding assistant locked inside an IDE. Instead, it executes tasks, invokes external tools, and maintains context over multiple sessions. In this context, an agent refers to a model that plans an approach, takes action, and reviews results in a continuous cycle.
Hermes Desktop puts a graphical layer on top of that identical engine. There’s no need to open a terminal window anymore. The interface displays responses in real time as they stream in, along with live updates on tool operations. A sidebar panel lets you preview websites, documents, and tool-generated results. On top of that, the app offers a built-in file browser, support for voice input and output, and a configuration panel.
Because sessions sync across all surfaces, you can kick off a task inside the desktop app, then pick it up right where you left off using the CLI or TUI, and vice versa. Since the underlying state is shared rather than copied, everything stays consistent.
For macOS and Windows users, dedicated installers are available. Linux users can grab it from the terminal on any distro using an install script with an --include-desktop parameter, which builds it alongside an existing setup.
The Closed Learning Loop
The Nous Research team says Hermes features a closed learning loop, the key detail that sets it apart from being just another chat interface. After completing a complex task, the agent generates a reusable skill, and those skills get refined automatically as they are used again.
Memory stays persistent and is managed by the agent itself, with reminders to save important information at intervals. When recalling information across different sessions, the system uses FTS5 search combined with LLM-based summarization. User profiling happens through Honcho’s dialectic user modeling approach. The longer you work with Hermes, the more context it retains and reuses. Skills adhere to the agentskills.io open specification.
How It Connects, Schedules, and Sandboxes
Hermes links up with multiple messaging platforms through a single gateway. The supported platforms visible in the desktop app include Telegram, Discord, Slack, WhatsApp, Signal, Email, and CLI. This means you can kick off a task on one platform and seamlessly continue it on another.
For scheduling, you can set up recurring reports, backups, and briefings using plain-language instructions. These run automatically in the background on the gateway’s built-in cron scheduler.
Delegation lets the agent spin up isolated subagents, each with their own conversation thread and terminal environment. Each subagent acts as an independent worker focused on a single job. Python RPC scripts streamline complicated multi-step workflows into turn-based interactions with zero added context cost.
Execution happens inside sandboxes. The desktop offers five backend options: local, Docker, SSH, Singularity, and Modal. The system enforces container-level hardening and namespace isolation, which restricts what a running process can access or modify.
Out-of-the-box tools cover web search, browser automation, image recognition, image creation, text-to-speech, and multi-model reasoning. Hermes also integrates with third-party tools through MCP, which stands for Model Context Protocol, a framework designed for connecting external tools.
Hermes is provider-agnostic, so supplying your own API keys is entirely optional. Nous Portal packages everything under a single subscription. Subscription levels include Free, Plus, Super, and Ultra. Paid plans come with monthly credit allowances and access to over 300 models, plus built-in tool functionality.
The Tool Gateway routes several features through one shared account. Web search leverages Firecrawl, image generation relies on FAL, text-to-speech uses OpenAI, and the cloud browser is powered by Browser Use.
Strengths and Questions
Strengths:
- Standalone installers eliminate the need to touch a terminal for most users
- Live streaming output and inline previews let you easily inspect what tools are doing
- Persistent memory and self-improving skills cut down on repetitive instruction overhead
- Provider-agnostic architecture keeps you from being locked into one service
- The MIT license lets you audit the code, self-host it, and make your own modifications
Questions:
- Since this is a public preview, rough edges and bugs are expected
- Autonomous memory management and scheduling raise concerns about proper oversight
- The Linux desktop version still requires terminal interaction for installation
- The wide range of capabilities may present a steeper learning curve for newcomers
Key Takeaways
- Nous Research has put Hermes Desktop into public preview, a native app spanning macOS, Windows, and Linux for its open-source Hermes Agent.
- The interface shares one unified agent core, along with configuration, API keys, sessions, skills, and memory with the CLI and gateway, so conversations carry over seamlessly between surfaces.
- It runs without a terminal, offering streaming tool output, a side-by-side preview panel, a file browser, voice input and output, and a settings interface.
- Hermes works with any provider and carries an MIT license, supporting Nous Portal, OpenRouter, OpenAI, or any compatible endpoint.
- The current release is Hermes Agent v0.15.2, powered by a closed learning loop, MCP tool integration, and five sandbox execution backends.
Check out the Project here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and subscribe to our Newsletter. Wait! Are you on Telegram? Now you can join us on Telegram as well.
Interested in partnering with us to promote your GitHub repo, Hugging Face page, product launch, or webinar? Connect with us
Michal Sutter is a data science expert who holds a Master of Science in Data Science from the University of Padova. He has strong skills in statistical analysis, machine learning, and data engineering, and is skilled at turning complicated data into useful insights.
