What Is the Hermes Agent?

Hermes in One Paragraph

Hermes Agent is a free, open-source AI agent framework created by Nous Research. It runs as a service — locally or on a server — and turns any supported AI model into a persistent agent you can interact with through messaging platforms, a web dashboard, or a standards-compatible REST API. Hermes supports multiple model providers, has built-in memory, can call real-world tools (web search, shell commands, image generation), and ships with a skill system for extending its capabilities. Because it is open source, you can inspect, modify, and self-host every part of it.

What Hermes Is Designed For

Based on the upstream source, Hermes is well suited for:

• Running a persistent AI agent on messaging platforms — connect Telegram, Discord, Slack, or other channels so the agent is always available without opening a browser tab.
• Exposing an OpenAI-compatible API endpoint — serve any compatible model behind a standard /v1/chat/completions interface that existing tools and libraries can call without modification.
• Agentic research and web tasks — have the agent search the web via Tavily, Exa, Parallel, or Firecrawl and synthesize results on your behalf.
• Automated shell and code execution — run terminal commands and scripts with configurable approval policies so you stay in control of what executes.
• Image generation workflows — produce images from prompts using FAL.ai without switching to a separate tool.

Key Capabilities

These capabilities are confirmed present in Hermes v2026.4.16 based on the upstream source code:

OpenAI-Compatible API Server

Hermes runs a standards-compatible API server on port 8642, enabled via the API_SERVER_ENABLED=true environment variable. It supports the standard endpoints — /v1/chat/completions, /v1/models, /v1/responses, and /v1/runs with server-sent event streaming. Session continuity across requests is maintained via previous_response_id, named conversation identifiers, or the Runs API — see the upstream API server docs for details. Any tool that speaks the OpenAI API format can point at Hermes directly.

Image Generation

The agent can generate images on demand using FAL.ai as the backend. Multiple models are supported — including FLUX variants, GPT Image, Ideogram, and Recraft. You supply a FAL API key (FAL_KEY), choose a default model, and the agent handles generation from natural-language prompts. Nous Research also operates a managed gateway mode for subscribers without their own FAL key.

Web Search

Hermes supports four first-class search providers: Tavily, Exa, Parallel, and Firecrawl. The agent auto-selects whichever provider key is configured. A free DuckDuckGo search skill is also available as an optional add-on for users who have not set up a paid search key.

Shell Execution with Approval Controls

The terminal tool can run commands locally, inside Docker, over SSH, or in Modal/Singularity/Daytona environments. An approval policy (approvals.mode) lets you set the agent to “manual” (confirm every command), “smart” (confirm risky ones), or “off” (fully autonomous). Per-command allowlists control which commands auto-approve for a session or permanently.

Vision and Audio Input

Hermes routes vision inputs through a centralized auxiliary router that supports OpenRouter, Nous, Codex, native Anthropic, and custom OpenAI-compatible endpoints. Audio transcription is available via a dedicated transcription tool.

Memory

The agent maintains session continuity and persistent memory across interactions, so context carries forward between conversations without you needing to re-explain your preferences or prior work.

Dashboard

The hermes dashboard command starts a separate local web server on port 9119 (bound to localhost only, no built-in auth required). It gives you config editing, log viewing, session history, skill management, cron scheduling, and environment variable control — all without touching config files by hand.

Skills Hub: 652 Skills Across 4 Registries

Hermes ships with a core set of built-in capabilities and extends them through the Skills Hub — 652 skills across four registries: 72 built-in, 59 optional, and 521 community. The Hub includes media generation skills: video (ascii-video, manim-video), music and audio generation (heartmula, audiocraft-audio-generation), and image generation (pixel-art, p5js) alongside the FAL.ai built-in. MCP servers are also supported — add them via a mcp_servers: block in your config.yaml or via the hermes mcp CLI command.

Hermes vs OpenClaw vs Other Agent Frameworks

Hermes and OpenClaw are both open-source AI agent frameworks — they share a broad goal but come from different organizations and have different design emphases. Hermes is built by Nous Research and leans toward developer-facing capabilities: a standards-compatible API server, strong model provider flexibility, and shell execution tooling. OpenClaw has a larger community skill ecosystem and broader messaging platform coverage. Neither is a reduced variant of the other.

For a full side-by-side breakdown, see the Hermes vs OpenClaw comparison page.

How to Try Hermes

Option 1: Install It Yourself

Clone the upstream repo and follow the self-host instructions. You will need Docker, a model provider API key, and some familiarity with YAML config files. Plan on an hour or more for initial setup.

Read the Hermes install guide →

Option 2: Deploy via OpenClaw Launch

OpenClaw Launch hosts Hermes for you — server, Docker, networking, SSL, and updates all handled. Pick your model, connect your messaging platform, and your agent is live in seconds. No config files, no server management.

Where Hermes Lives

The canonical source is github.com/NousResearch/hermes-agent. That is the ground truth for what the current version supports — including config field names, environment variables, and platform options. The README and source files there are more reliable than any third-party summary, including this one.

Feature Comparison

Frequently Asked Questions

Is Hermes free?

Yes. The Hermes Agent source code is free and open source. You can clone it and run it at no cost on your own hardware. You will still pay for the AI model API calls you make (billed by whatever provider you configure), and if you run it on a VPS you pay for that server. Managed hosting services like OpenClaw Launch charge a monthly fee that covers the server so you do not have to manage one yourself.

Who makes Hermes?

Hermes Agent is built and maintained by Nous Research, an AI research organization known for its open-weight model work. The agent framework is separate from Nous's model releases (also named “Hermes”) — the agent is what runs and orchestrates the model, not the model itself.

Does Hermes run offline or without an internet connection?

The Hermes server process itself can run on a local machine with no public internet exposure. However, it needs to reach your configured model provider API to generate responses. If you point Hermes at a locally running model (e.g., via a local OpenAI-compatible server), the full stack can operate without any outbound internet connection.

Can Hermes call my own API or use my own model?

Yes. Hermes accepts any provider that exposes an OpenAI-compatible endpoint. You can point it at a local model server, a custom API gateway, or any of the supported cloud providers by setting the appropriate environment variables in your config.

What does Hermes cost to run in practice?

Infrastructure cost varies by how you host it. A minimal VPS for the Hermes container typically runs $5–10/month. Model API cost depends entirely on usage — light personal use is usually a few dollars a month; heavier automated workflows cost more. The Hermes software itself adds no per-request fee on top of the provider cost.

Related Guides

• How to install Hermes Agent — Step-by-step self-host setup
• Hermes Agent on GitHub — Navigating the upstream repository
• Hermes Agent memory configuration — How persistent memory works
• Hermes vs OpenClaw — full comparison — Side-by-side breakdown
• Deploy Hermes Agent in seconds — Using OpenClaw Launch managed hosting