An open-source AI agent developed by Nous Research. It runs persistently on your own machine or on a VPS costing a few dollars a month, and you can talk to the same agent through Telegram, Discord, Slack, WhatsApp, Signal, email, or the CLI. Its defining trait is a “gets smarter the more you use it” design: persistent memory that survives across sessions, plus a learning loop that automatically turns experience from complex tasks into reusable skills. It ships with more than 40 built-in tools, including browser automation, web search, voice, image generation, MCP integration, and subagent delegation. The software itself is free under the MIT license — your only cost is API usage from whichever LLM provider you choose.
Key Features
- Persistent memory so context never disappears: The agent curates and retains the premises of your conversations and projects, and references them across sessions. It can also search the full text of past sessions and surface summaries, so “that thing I asked you last week” actually works
- A learning loop that generates skills from experience: Once it completes a complex task, it writes the procedure out as a skill and refines it on subsequent uses. Skills follow the open agentskills.io standard, so externally authored skills can be imported as well
- 15+ platforms through a single gateway: Telegram, Discord, Slack, WhatsApp, Signal, email, and the CLI are all handled by one gateway process. Voice message transcription is supported, and conversations continue no matter which entry point you use
- Sandboxed execution and flexible deployment targets: Choose from multiple terminal backends — local, Docker, SSH, Singularity, Modal, Daytona, and Vercel. It runs on a $5/month VPS, and with Modal or Daytona it can hibernate while idle to keep costs down
- Browser automation, scheduled tasks, and subagents: Web search and browser control, cron tasks configured in natural language (automating recurring reports or backups), and subagent spawning for parallel workstreams all come standard
- MCP integration and freedom to switch models: Connect to external tools via the Model Context Protocol. Models can come from Nous Portal, OpenRouter, OpenAI, Anthropic, Google, or local models, and the
hermes modelcommand switches between them — no vendor lock-in
Pricing
Hermes Agent itself is free, open-source software under the MIT license. Your actual cost depends on which LLM provider you use. If you go with Nous Portal from the developer, a single subscription covers model access plus the built-in tools (web search, scraping, image generation, browser control, and speech).
| Plan | Monthly price | What’s included |
|---|---|---|
| Hermes Agent itself | $0 | MIT license. All features available. You cover only LLM API usage |
| Nous Portal Free | $0 | Pay-as-you-go model access |
| Nous Portal Plus | $20 | Monthly credits, hosted tool access, raised rate limits |
| Nous Portal Super | $100 | More monthly credits and higher rate limits than Plus |
| Nous Portal Ultra | $200 | Top-tier monthly credits and rate limits |
Using a local LLM brings API costs to zero (though you will need adequate GPU memory). Configurations using your own OpenRouter or Anthropic API keys work fine too.
Pricing is current as of August 2026. The credit amounts and rollover caps included in each Nous Portal plan are subject to change, so please check the official site and Nous Portal for the latest pricing.
Pros & Cons
✅ Pros
- Free and open source (MIT license), so you can read the source to verify behavior or extend it yourself
- Persistent memory and skill learning mean you stop repeating the same instructions — it fits your workflow better the longer you use it
- Accessible from the chat apps you already use, so there is no separate interface to open
- You are free to pick your model provider, including local LLMs, and to switch later. API costs stay under your control
- Can run persistently on a $5-class VPS, with cron tasks handling recurring work
⚠️ Cons
- Self-hosting is assumed, so installation, API key configuration, and keeping the process running are on you (though one-line installers are provided)
- Not entirely free in practice: LLM API usage costs money, and browser control or long autonomous runs make token consumption hard to predict
- Running local models requires a machine with a reasonable amount of GPU memory
- Because it is an agent that executes commands autonomously, poor permission design or a badly chosen sandbox is dangerous. You need to understand the sandboxed execution options
- Development moves fast, and configuration or behavior can change between versions
Comparison with Similar Services
| Criteria | Hermes Agent | OpenClaw | Claude Code | OpenHands |
|---|---|---|---|---|
| Delivery model | Self-hosted (OSS, MIT) | Self-hosted (OSS) | Terminal-resident CLI (commercial) | Self-hosted (OSS) |
| Primary use | General-purpose always-on assistant | General-purpose always-on assistant | Delegating coding and development work | Autonomous software development |
| Main entry points | Telegram, Discord, Slack, WhatsApp, Signal, email, CLI | Mainly chat apps | Terminal | Web UI, CLI |
| Long-term memory | Persistent memory plus full-text session search | Yes | Per-project context management | Limited |
| Automatic skill generation | Yes (generated from experience, self-improving) | Extensions are mainly plugin-based | Agent Skills (typically authored manually) | Limited |
| Model choice | Any provider, local LLMs supported | Any provider | Anthropic models | Any provider |
Who Is It For
- People who want their own AI assistant, callable from chat apps, running persistently on their own machine or a VPS
- Anyone tired of re-explaining the same background every time and looking for an agent that remembers context
- People who want to automate recurring work — periodic reports, backups, information gathering — using natural language
- Those who want to control where their data lives and which model they use, without depending on a cloud service
- Owners of local LLMs who want to run an agent without paying for API usage
- Developers who want to read an OSS agent implementation and learn how memory and skill learning are built
Summary
Hermes Agent is an open-source AI agent from Nous Research built around a “continuously gets smarter” design centered on persistent memory and automatic skill generation. Its strengths are the convenience of being callable from everyday chat apps combined with the flexibility to choose both your model and your deployment target. The flip side is that self-hosting is assumed, so setup and permission design are your responsibility. A good starting point is installing it locally or on an inexpensive VPS, handing it small tasks using an API key you already have, and seeing how its memory and skills develop.