Hermes Agent — Summary

Hermes Agent by NousResearch is a production-grade self-improving AI agent with 167,908 stars — the second most-starred framework in this batch. It is a standalone Python CLI application (hermes binary) with a full TUI (prompt_toolkit), 7 terminal backends (local/Docker/SSH/Singularity/Modal/Daytona/Vercel Sandbox), a messaging gateway (Telegram/Discord/Slack/WhatsApp/Signal), a web dashboard (React/Vite), a built-in cron scheduler, multi-provider LLM support (200+ via OpenRouter), FTS5 SQLite session search, and a closed learning loop that creates skills from experience and improves them during use. Skills follow the agentskills.io open standard. The "self-improving" claim is implemented via autonomous skill creation after complex tasks, skill self-improvement during use, periodic memory nudges, and dialectic user modeling via Honcho. Seven terminal backends make it uniquely deployable beyond the laptop — on VPS, GPU cluster, or serverless (Modal/Daytona/Vercel Sandbox). Differs from seeds: no seed is a self-hosted personal agent of this scope. Closest to claude-flow (Archetype 3: MCP-anchored toolserver, continuous execution) but Hermes is provider-agnostic and ships as a Python package rather than npm. The learning loop (skills from experience) is unique — no seed implements autonomous skill authoring.

Hermes Agent — Overview

Origin

GitHub: https://github.com/NousResearch/hermes-agent
Stars: 167,908
License: MIT
Language: Python (primary), TypeScript (web dashboard)
Version: 0.14.0 (from pyproject.toml)
Last commit: 2026-05-26
Organization: Nous Research (AI research company)

Philosophy

From README:

"The self-improving AI agent built by Nous Research. It's the only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, searches its own past conversations, and builds a deepening model of who you are across sessions."

"Run it on a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle. It's not tied to your laptop — talk to it from Telegram while it works on a cloud VM."

"Use any model you want — Nous Portal, OpenRouter (200+ models), NovitaAI, NVIDIA NIM, Xiaomi MiMo, z.ai/GLM, Kimi/Moonshot, MiniMax, Hugging Face, OpenAI, or your own endpoint."

Key Design Decisions

1. Learning loop: skills created from task experience, improved during use — not static library.
2. Not tied to laptop: 7 terminal backends including serverless (Modal/Daytona) and SSH.
3. OpenClaw migration support: explicit migration path from the predecessor project.
4. agentskills.io compatibility: skills follow an open standard for portability.
5. Exact version pinning: all dependencies pinned to exact versions (supply chain security response to Mini Shai-Hulud worm incident, 2026-05-12).

Hermes Agent — Architecture

Distribution

# Linux/macOS/WSL2
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
source ~/.bashrc
hermes

# Windows PowerShell
iex (irm https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.ps1)

Package

• PyPI: hermes-agent (pip installable)
• Package: pyproject.toml with [project.scripts]
  • hermes → hermes_cli.main:main
  • hermes-agent → run_agent:main
  • hermes-acp → acp_adapter.entry:main

Directory Tree (selected)

hermes-agent/
├── hermes/                  # Core library
├── hermes_cli/              # CLI and TUI
├── agent/                   # Agent loop, model tools
├── gateway/                 # Messaging platform gateway
├── tui_gateway/             # TUI gateway
├── web/                     # React/Vite web dashboard
├── skills/                  # Skill categories
│   ├── software-development/  # plan, TDD, debugging, etc.
│   ├── devops/
│   ├── mlops/
│   ├── research/
│   ├── social-media/
│   └── ...24 categories total
├── acp_adapter/             # ACP protocol adapter
├── acp_registry/            # ACP registry
├── providers/               # LLM provider adapters
├── mcp_serve.py             # MCP server
├── hermes_state.py          # SQLite state store
├── trajectory_compressor.py # Training data compression
├── pyproject.toml           # v0.14.0
└── uv.lock                  # Exact dependency pins

Required Runtime

• Python 3.11
• uv (package manager, recommended) or pip
• Node.js (for web dashboard)
• Provider API key or Nous Portal subscription

Terminal Backends (7)

Local, Docker, SSH, Singularity, Modal (serverless), Daytona (serverless), Vercel Sandbox

Hermes Agent — Components

CLI Subcommands

Skills (24 categories, 100+ individual skills)

Messaging Gateway

Platforms: Telegram, Discord, Slack, WhatsApp, Signal, Email, Home Assistant

MCP Integration

• mcp_serve.py — bundled MCP server
• HTTP transport with auto-reconnect
• JSON Schema type inference for tool inputs

ACP (Agent Communication Protocol)

• acp_adapter/ — ACP protocol adapter
• acp_registry/ — ACP registry

Web Dashboard

• web/ — React + Vite
• Stack: @nous-research/ui, xterm.js, observable/plot, react-three/fiber

State Store

• hermes_state.py — SQLite with WAL mode, FTS5 full-text search, session history

Trajectory Compression

• trajectory_compressor.py — compresses sessions for training the next generation of tool-calling models

Hermes Agent — Uniqueness

differs_from_seeds

No seed is a self-improving personal agent of this scope. Closest to claude-flow (continuous execution, SQLite memory, multi-agent) but claude-flow wraps Claude Code while Hermes is a standalone Python runtime that works with any provider. The learning loop (autonomous skill creation from experience + skill self-improvement) is the single most distinctive feature — no seed implements this. The 7 terminal backends (including serverless Modal/Daytona) enable deployment patterns no other framework in the catalog supports. The trajectory compression feature (converting sessions into LLM training data) positions Hermes as both an agent framework AND a research platform for training better tool-calling models.

Most Distinctive Features

1. Autonomous skill creation: skills authored from task experience, not just pre-defined.
2. 7 terminal backends: runs on VPS, Docker, SSH, Modal (serverless), Daytona — the only framework designed to run off-device.
3. Trajectory compression: converts sessions into training data — dual-purpose as research tool.
4. agentskills.io standard: skills follow an open standard for portability across compatible frameworks.
5. Supply chain security: exact dependency pinning in response to the 2026-05-12 Mini Shai-Hulud worm incident.

Observable Failure Modes

1. Complexity: many features, many providers, many backends — higher configuration burden than simpler harnesses.
2. Exact pinning cost: every update requires manually bumping pins — faster development but higher maintenance.
3. Nous Portal coupling: the full tool gateway experience requires a Nous Portal subscription.
4. Learning loop quality: autonomous skill creation quality depends on the quality of task execution — bad habits get codified.