Hermes One

Hermes One is an MIT-licensed desktop AI agent framework built around a self-improvement loop: the agent writes and refines its own skills during use rather than relying on a fixed toolset. Out of the box it ships with 14 toolsets covering web search, code execution, image generation, file ops, and scheduled automations driven by natural-language cron-like instructions, plus a memory layer with full-text search and LLM-mediated cross-session recall.

It's aimed at developers, researchers, and power users who want an autonomous agent they can pipe into the channels they already live in - Telegram, Discord, Slack, WhatsApp, Signal, or a local CLI - rather than a hosted chat UI. Model routing is BYO: OpenRouter (300+ models), OpenAI, Anthropic, Gemini, xAI, NVIDIA NIM, Hugging Face, Ollama, and LM Studio are all supported, so you pay only the underlying inference bill. The repo has ~12.9k GitHub stars, which suggests an active community but a moving target.

The parallel-subagent spawning model and Nous Portal integration position it more as a research-grade agent harness than a polished consumer product. Expect to wire up your own API keys, host it yourself, and handle the operational side of long-running agents.