Skip to content
Roman's notes

Hermes Agent

Hermes Agent is a self-hosted autonomous agent runtime built by Nous Research and released under the MIT license. It is a long-running process that maintains its own persistent memory and skill library across sessions, exposes the agent through messaging gateways and a CLI, and orchestrates tool calls, subagents, and scheduled tasks on the user’s own infrastructure rather than a vendor cloud.

🔑 Key capabilities

Section titled “🔑 Key capabilities”
  • Persistent memory — agent-curated knowledge base with full-text search (FTS5) and cross-session recall via LLM summarization
  • Auto-generated skills — the agent creates reusable skills during operation using the open Agent Skills format; community-contributed skills are available via agentskills.io
  • Subagents — isolated subagents with their own conversations, terminals, and Python RPC scripts for parallel, zero-context-cost pipelines
  • Built-in scheduling — natural language cron jobs for reports, backups, and briefings running unattended
  • 47 built-in tools — web search, browser automation, vision, image generation, text-to-speech, and more
  • MCP integration — connects to any Model Context Protocol server with tool filtering

🌐 Messaging gateways

Section titled “🌐 Messaging gateways”

Supports 14+ platforms: CLI, Telegram, Discord, Slack, WhatsApp, Signal, Matrix, Mattermost, Email, SMS, DingTalk, Feishu, WeCom, Home Assistant. Conversations are portable — start on one platform, continue on another.

🛠️ Deployment backends

Section titled “🛠️ Deployment backends”
BackendNotes
LocalLinux, macOS, WSL2
DockerContainer hardening and namespace isolation
SSHRemote execution
SingularityHPC environments
Modal / DaytonaServerless, near-zero cost when idle

💻 Desktop app

Section titled “💻 Desktop app”

Hermes Desktop is a native installer that bundles the agent runtime with a GUI.

  • macOS 12+ — DMG installer
  • Windows 10/11 — EXE installer
  • Linux — curl installer (terminal-based)

The desktop app exposes the same capabilities as the headless runtime — persistent memory, auto-generated skills, subagents, scheduling, 47+ built-in tools, MCP — and adds direct integration with Nous Portal (subscription tiers with monthly credits and access to 300+ models). Sandboxing offers five backends: local, Docker, SSH, Singularity, Modal.

🤖 Model support

Section titled “🤖 Model support”

Compatible with Nous Portal, OpenRouter, OpenAI, and custom endpoints.

🧪 Research features

Section titled “🧪 Research features”
  • Trajectory export for research and evaluation
  • RL training pipeline via the Atropos framework
  • Honcho-based dialectic user modeling

Sources

Section titled “Sources”