Hermes Agents

1. Overview

Hermes Agent is an open-source (MIT) self-hosted AI agent built by Nous Research — the agent that grows with you. It is neither a coding copilot tethered to an IDE, nor a chatbot wrapper around a single API. It runs on your own server, remembers what it learns, and becomes more capable the longer it runs. With GravitexAI (OpenAI-compatible API) one key gives Hermes access to GPT, Claude, Gemini, DeepSeek and 400+ other leading models.

Persistent Memory

Remembers projects and solutions across sessions

Auto Skill Synthesis

Generates reusable Skills from experience

Multi-platform Reach

Telegram / Discord / Slack / WhatsApp / Signal / Email / CLI

Real Sandboxing

Local / Docker / SSH / Singularity / Modal backends

Website: hermes-agent.nousresearch.com ｜ Repo: github.com/NousResearch/hermes-agent ｜ License: MIT

2. Core Capabilities

Persistent memory — local vector store retains conversations and project knowledge
Learning loop — successful solutions are crystallised into reusable Skills
Scheduled automations — describe cron jobs in natural language
Delegated subagents — isolated terminals & Python RPC, no context pollution
Full web control — browser automation, vision, image generation, TTS
Real sandboxes — local, Docker, SSH, Singularity, Modal

3. Prerequisites

Create an API key in the GravitexAI console
A server or workstation: Linux / macOS / WSL2 / Android (Termux)
Windows users: PowerShell (early beta)

Hermes Agent requires a model with ≥ 64K tokens of context. Claude, GPT, Gemini and DeepSeek all qualify.

4. Install

One command handles everything:

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

irm https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.ps1 | iex

All dependencies are placed under ~/.hermes/hermes-agent. No manual pip install needed.

5. Configure GravitexAI

Hermes Agent connects to any OpenAI-compatible service via its Custom Endpoint path.

Option A — Interactive CLI (recommended)

hermes model

Pick Custom endpoint (self-hosted / VLLM / etc.) and fill in:

Field	Value
API Base URL	`https://api.gravitex.ai/v1`
API Key	Your GravitexAI key (`sk-...`)
Model name	`claude-sonnet-4-6` (or any GravitexAI model ID)

Settings persist to ~/.hermes/config.yaml.

Option B — Edit config.yaml manually

Open ~/.hermes/config.yaml and add:

model:
  provider: custom
  model: claude-sonnet-4-6
  base_url: https://api.gravitex.ai/v1
  api_key: sk-your-gravitexai-key

When base_url is set, Hermes ignores provider and calls the endpoint directly. The API key is read from model.api_key, falling back to the OPENAI_API_KEY environment variable.

Option C — Environment variables

export OPENAI_BASE_URL=https://api.gravitex.ai/v1
export OPENAI_API_KEY=sk-your-gravitexai-key
hermes model   # the Custom Endpoint wizard will pick them up

Option D — Multiple named endpoints

Declare several endpoints in advance and switch at runtime:

custom_providers:
  - name: gravitex
    base_url: https://api.gravitex.ai/v1
    key_env: GRAVITEX_API_KEY
    api_mode: chat_completions
  - name: gravitex-anthropic
    base_url: https://api.gravitex.ai/anthropic
    key_env: GRAVITEX_API_KEY
    api_mode: anthropic_messages

Switch with hermes model, or set model.provider: gravitex directly.

6. Run & Verify

hermes setup     # first-run wizard (skip if already configured)
hermes doctor    # health check: validates key & /models
hermes start     # launch the gateway

You can now chat with Hermes from any connected channel — CLI, Telegram, Discord, Slack, etc.

7. Recommended Models

Use case	Model ID	Notes
Complex tasks	`claude-sonnet-4-6`	Best tool calling & long tasks
Coding	`deepseek-v3-2-251201`	Great code, great price
Daily chat	`gpt-5.5`	Strong all-rounder
Long context	`gemini-3.1-pro-preview`	Massive context window
Low latency	`claude-haiku-4-5-20251001`	Fast responses
Heavy reasoning	`claude-opus-4-7`	Complex reasoning, algorithms / logic

Browse all models in the GravitexAI marketplace.

8. Messaging Channels

Platform	Config location
Telegram	`channels.telegram.token`
Discord	`channels.discord.bot_token`
Slack	`channels.slack.bot_token`
WhatsApp	QR scan
Signal	linked-device QR
Email	SMTP / IMAP
CLI	`hermes chat`

9. Extensions

Skills

hermes skills search kubernetes
hermes skills install openai/skills/k8s

Hermes automatically synthesises Skills from your sessions and reuses them later — the feature that sets it apart from most agent frameworks.

MCP servers

Add MCP tools to ~/.hermes/config.yaml:

mcp_servers:
  github:
    command: npx
    args: ["-y", "@modelcontextprotocol/server-github"]
    env:
      GITHUB_PERSONAL_ACCESS_TOKEN: "ghp_xxx"

Sandbox backends

sandbox:
  backend: docker   # local / docker / ssh / singularity / modal

10. Commands

Command	Purpose
`hermes setup`	First-run wizard
`hermes model`	Choose / switch provider & model
`hermes doctor`	Health check
`hermes start` / `stop` / `restart`	Control the gateway
`hermes chat`	Terminal chat
`hermes skills`	Manage skills
`hermes logs --follow`	Tail logs
`hermes update`	Upgrade

11. Troubleshooting

401 / invalid API key

Verify api_key in ~/.hermes/config.yaml has no stray characters
When several *_API_KEY env vars exist, remember Hermes scopes each key to its own base URL

404 / endpoint not found

chat_completions mode: base_url must end with /v1
For native Claude protocol use https://api.gravitex.ai/anthropic with api_mode: anthropic_messages

Context too short

Hermes requires ≥ 64K context. Use claude-sonnet-4-6, gpt-5.4, gemini-3.1-pro-preview, etc.

Config changes ignored

config.yaml is the single source of truth. If old OPENAI_BASE_URL lingers in .env, run hermes setup to trigger the migration.

Chat AI

Coding

AI Agent

Tech Engineering

Translation

1. Overview

Persistent Memory

Auto Skill Synthesis

Multi-platform Reach

Real Sandboxing

2. Core Capabilities

3. Prerequisites

4. Install

5. Configure GravitexAI

Option A — Interactive CLI (recommended)

Option B — Edit config.yaml manually

Option C — Environment variables

Option D — Multiple named endpoints

6. Run & Verify

7. Recommended Models

8. Messaging Channels

9. Extensions

Skills

MCP servers

Sandbox backends

10. Commands

11. Troubleshooting

​1. Overview

Persistent Memory

Auto Skill Synthesis

Multi-platform Reach

Real Sandboxing

​2. Core Capabilities

​3. Prerequisites

​4. Install

​5. Configure GravitexAI

​Option A — Interactive CLI (recommended)

​Option B — Edit config.yaml manually

​Option C — Environment variables

​Option D — Multiple named endpoints

​6. Run & Verify

​7. Recommended Models

​8. Messaging Channels

​9. Extensions

​Skills

​MCP servers

​Sandbox backends

​10. Commands

​11. Troubleshooting

1. Overview

2. Core Capabilities

3. Prerequisites

4. Install

5. Configure GravitexAI

Option A — Interactive CLI (recommended)

Option B — Edit config.yaml manually

Option C — Environment variables

Option D — Multiple named endpoints

6. Run & Verify

7. Recommended Models

8. Messaging Channels

9. Extensions

Skills

MCP servers

Sandbox backends

10. Commands

11. Troubleshooting