Direct answer: Hermes AgentAI agentAn AI program that performs work steps by itself, for example reading a message, drafting a reply, and recording the result.Open the glossary is an open-source AI agent (MIT license) built by Nous Research that you install on your own computer or server. It remembers work across sessions, writes its own skills, runs terminal commands, reads and writes files, drives a browser, and answers from Telegram, Discord, Slack, WhatsApp, Signal, email, or the terminal. The software is free. You pay only the model provider, for example Nous Portal from $20 per month or any other APIAPIThe official door 2 systems use to exchange data, without anybody copying it by hand.Open the glossary you choose.
Main condition: Hermes Agent is worth installing when a repeated task needs tools, not only answers, and 1 person owns the server, the approvals, and the model bill. Limit: Hermes does not replace the ChatGPT or ClaudeClaude ProThe paid subscription for the Claude AI assistant from Anthropic. It unlocks connectors to outside tools.Open the glossary app for personal questions, and it is not a sandbox. Its terminal tool runs as the same OS user. Real isolation comes only from the docker or ssh backend.
We read the official Hermes Agent documentation, the Nous Portal plans page, and the GitHub repository on 19 September 2026. On the same day we installed Hermes Agent v0.21.3 in an Ubuntu 24.04 container with 1 command, then ran hermes doctor, hermes cron create, and the hermes whatsapp wizard. Every terminal screenshot in this article comes from that test. Phone numbers and host names are masked.
The problem Hermes Agent solves
A normal AI chat stops at the chat window. You ask, it answers, and you copy the answer somewhere else. A task with 5 steps, such as reading 3 sources, summarising them, writing a file, and sending it to the team group, stays your job. Tomorrow you repeat it.
Hermes Agent moves that work to a server. It keeps memory in MEMORY.md and USER.md, stores procedures in SKILL.md files, and runs scheduled tasks through cron while your laptop is closed. The Indonesian market has started to search for this tool. We described the shift of interest from OpenClaw to Hermes in our analysis of OpenClaw and Hermes Agent trends in Indonesia (Indonesian). In our Search Console the query "apa itu hermes agent" appeared without a page that answers it. This article is that page.
How Hermes Agent works
Hermes Agent has 1 core and 3 entry points. The diagram below maps the flow from your prompt to the model answer.
Entry points: CLI, Desktop, and gateway
The hermes command opens a chat session in the terminal. hermes --tui opens a richer terminal interface. Hermes Desktop is a native app for macOS, Windows, and Linux that uses the same core, configuration, sessions, and memory as the CLI. The gateway connects the agent to Telegram, Discord, Slack, WhatsApp, Signal, email, Matrix, and other platforms. Source: Hermes Desktop documentation and Messaging Gateway documentation, 19 September 2026.
The core: memory, skills, and tools
Memory stores facts about you, your projects, and your environment. Skills store step-by-step procedures in SKILL.md files that follow the open agentskills.io standard. Tools are functions the agent can call: terminal, file editing, web search, browser, cron, delegation to sub-agents, and MCPMCPA standard that lets an AI program use your tools and data under clear permissions.Open the glossary servers. The agent also loads project context files such as AGENTS.md, CLAUDE.md, and .hermes.md automatically. Source: Features Overview. We cover memory and skills for operations in our article on Hermes Agent memory and skills for business operations (Indonesian).
Model provider: 1 login or your own API
Hermes works with any API that speaks the OpenAI format. The documentation lists OpenRouter, Nous Portal, OpenAI, Anthropic, Google Gemini, z.ai, Kimi, MiniMax, and local models through Ollama, vLLM, llama.cpp, or SGLang. Nous Portal is the vendor-recommended path: 1 OAuth login gives access to the model catalog and the Tool Gateway for web search, image generation, TTS, and a cloud browser. The command is hermes setup --portal. Source: Hermes Agent FAQ and the Nous Portal page.
Security defaults: approvals and the blocklist
A fresh install uses approvals.mode: smart. Low-risk commands are approved automatically, dangerous commands are denied, and uncertain cases escalate to a manual prompt. A prompt with no answer within 300 seconds is denied. A hardline blocklist refuses commands such as rm -rf / and cannot be switched off. The file-write tools are blocked for ~/.ssh/, ~/.aws/, and .env files. The gateway denies every user who is not allowlisted or paired. Source: the guide to running Hermes on a work machine. In our test, hermes config get approvals.mode returned smart and terminal.backend returned local.
3 ways to use Hermes Agent
This table compares the 3 entry points as the Hermes Agent documentation described them on 19 September 2026.
| Item | CLI and TUI | Hermes Desktop | Messaging gateway |
|---|---|---|---|
| Fits | Server owners, developers, first tests | Daily users on a laptop | Teams that work in chat |
| How to start | hermes or hermes --tui | The installer from the official site or hermes desktop | hermes gateway setup then hermes gateway install |
| Platforms | Linux, macOS, WSL2, Windows, Termux | macOS Apple Silicon, Windows, Linux | A server or VPS that stays up 24 hours |
| Access for many people | No | No | Yes, with an allowlist or DM pairing |
| Scheduled tasks | Needs a running gateway | Needs a running gateway | Yes, cron and webhooks |
Pricing: free software, paid models
The official FAQ states that Hermes Agent is free and open source. You pay the API usage of the provider you choose. Local models cost nothing to run. Nous Portal sells monthly plans that we read on the Manage Subscription page on 19 September 2026. Prices are in US dollars and can change.
| Plan | Price per month | Monthly credits | Rollover cap | Page note |
|---|---|---|---|---|
| Free | $0 | $0 | None | Free models only, standard rate limits |
| Plus | $20 | $22 | $10 | 200+ models, hosted tool usage, high rate limits |
| Super | $100 | $110 | $50 | Same as Plus with more credits |
| Ultra | $200 | $220 | $100 | Same as Plus with more credits |
Source: the Nous Portal plans page. The Nous Portal privacy policy states that prompts, inputs, and outputs can be stored and used for training unless you enable Privacy Mode in Account Settings. When client data passes through Nous Portal, enable Privacy Mode first. Source: the Nous Portal privacy policy. Other paths, such as the Anthropic or OpenAI API directly, follow that provider's policy. Local models send data nowhere.
Prerequisites before the test
- 1 machine with Linux, macOS Apple Silicon, Windows 10 or 11, or WSL2. Tier 1 on the Platform Support page. Intel macOS, pip, and brew are not supported.
- Git installed. On Linux,
curlandxz-utilsinstalled too. - 1 model provider account: Nous Portal, OpenRouter, Anthropic, OpenAI, or a local Ollama server.
- 1 real task to test, for example a summary of 3 industry news items every morning.
- 1 person who owns the server and decides on approvals.
Step 1: Install with 1 command
Open a terminal on the test machine. Run the command from the official site:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bashThe installer checks Python 3.11 through uv, Node.js, ripgrep, and ffmpeg, then clones the repository to ~/.hermes/hermes-agent/ and puts the hermes command in ~/.local/bin/. Windows users run iex (irm https://hermes-agent.nousresearch.com/install.ps1) in PowerShell. Source: the Installation page.
Verify: run source ~/.bashrc then hermes --version. In our test the output was Hermes Agent v0.21.3 (2026.9.14) with install method git and status Up to date.
Step 2: Check health with hermes doctor
Run hermes doctor. The command checks security advisories, Python, SQLite, required packages, configuration files, logged-in providers, the folder structure, and external tools. A green line means ready. A yellow line means optional or action needed.
Verify: no red lines. The yellow provider warnings disappear after Step 3. The SQLite warning on Ubuntu 24.04 comes from the system package and does not block the test.
Step 3: Connect a model provider
Pick 1 of 2 paths. Path A: hermes setup --portal opens the browser for the Nous Portal login, picks a model, and turns on the Tool Gateway in 1 command. Path B: hermes model lists the providers, including OpenRouter, Anthropic, OpenAI, and a Custom endpoint for Ollama. API keys live in ~/.hermes/.env. For a server without a browser, the documentation has an OAuth guide through SSH port forwarding.
Verify: run hermes and send 1 real task, for example "Find 3 AI agent news items from today and summarise them in 5 sentences". The agent must call web search, not guess. Note the token count on the status line.
Step 4: Lock access before you open the gateway
Before you connect Telegram or WhatsApp, set 3 things. First, the approvals mode: smart for personal use, manual for a server with client data. Second, a deny list for commands that must never run, such as git push --force*. Third, a per-platform user allowlist, never *.
Verify: hermes pairing list must be empty before anyone pairs. After the gateway is up, send 1 message from a number or account that is not allowlisted. The bot must stay silent or send a pairing code, never an answer. The full gateway guide is in our article on Hermes Agent for Telegram, Discord, Slack, and WhatsApp (Indonesian).
A simulated day at a fictional online shop
This table is a simulation with dummy data. Toko Kain Nusantara runs Hermes Agent on a VPS with Nous Portal Plus, a Telegram gateway for 2 admins, and 1 morning cron job.
| Time | Event | What Hermes records | Output |
|---|---|---|---|
| 07:00 | The cron job "Morning briefing" runs | 1 new session, 4 web searches, 1 summary | An 8-line Telegram message to the admin group |
| 09:15 | An admin asks "which batik fabrics are below 10 in stock?" | The MCP tool reads the stock spreadsheet | A list of 6 SKUs with quantities |
| 09:20 | The admin asks "draft a restock message to the supplier" | The supplier-draft skill is reused | A text draft, not sent |
| 11:40 | Someone outside the allowlist messages the bot | Denied, a pairing code is sent | No content answer |
| 16:00 | The admin asks "delete every file in the export folder" | Smart approvals escalate to a manual prompt | The server owner denies it |
| 23:59 | The day ends | 3 sessions, about 40 tool calls | Used credits recorded in Nous Portal |
The numbers in this table are an illustration, not a customer result. What you copy is the pattern: 1 cron job, 1 allowlist, 1 approvals setting, and 1 person who reads the log.
Checklist before you decide
- Write down 1 repeated task with a clear input, tools, and output. Owner: the process owner. Evidence: a 1-paragraph task.
- Install Hermes on 1 test machine with the official command. Owner: the server owner. Evidence: the
hermes --versionoutput. - Run
hermes doctorwith no red lines. Owner: the server owner. Evidence: a screenshot. - Connect 1 provider and run the task 5 times in 5 days. Owner: the process owner. Evidence: 5 outputs and the token counts.
- Set approvals, the deny list, and the allowlist before the gateway. Owner: the server owner. Evidence:
hermes config get approvals.mode. - Enable Privacy Mode in Nous Portal when client data passes through it. Owner: the data owner. Evidence: the Privacy Settings page.
- Count 1 week of cost and compare it with the working hours saved. Owner: the business owner. Evidence: 1 line of numbers.
- Stop criterion: when the team does not use the 5-day output, or nobody wants to own the server, stop the test and use a normal chat app.
Hermes Agent compared with chat apps and OpenClaw
This table uses each product's official pages on 19 September 2026. It does not judge answer quality, because that quality follows the model you pick.
| Item you can verify | Hermes Agent | Chat apps (ChatGPT, Claude) | OpenClaw |
|---|---|---|---|
| Where it runs | Your server or computer | The vendor server | Your server or computer |
| Software cost | Free, MIT | Subscription per user | Free, MIT |
| Access from WhatsApp and Telegram | Yes, built-in gateway | Limited to vendor features | Yes, built-in gateway |
| Scheduled tasks on your own server | Yes, cron and webhooks | No | Yes, cron and hooks |
| Maker | Nous Research, a venture-funded company | OpenAI, Anthropic | OpenClaw Foundation, a non-profit |
| GitHub stars | 246,927 | Not applicable | 390,059 |
Sources: the Hermes Agent FAQ, the What is OpenClaw page, and the GitHub API on 19 September 2026. Rama Digital recommendation: choose Hermes Agent when your team wants 1 model subscription for every tool and the server owner is at home with Python. Choose a normal chat app when your need stops at questions and answers. We wrote the full comparison with OpenClaw in our Hermes Agent vs OpenClaw article.
Frequently asked questions
Is Hermes Agent free? Yes. The software is open source under the MIT license. You pay model usage to the provider you choose. Nous Portal started at $20 per month on 19 September 2026, and local models through Ollama are free.
Is my data sent to Nous Research? According to the official FAQ, API calls go only to the model provider you configure and Hermes sends no telemetry. When that provider is Nous Portal, its privacy policy applies and Privacy Mode must be enabled for sensitive data.
Can I use Hermes Agent through WhatsApp? Yes, through 2 paths. The Baileys bridge uses a QR code from a dedicated number and is not official from Meta. The WhatsApp Business Cloud API adapter uses a Meta Business account and a public HTTPS webhookWebhookAn automatic message one system sends to another the moment something happens, for example an incoming WhatsApp message.Open the glossary.
What is the difference between memory and skills? Memory stores facts about you, your projects, and your preferences. Skills store step-by-step procedures. Both persist across sessions and live in ~/.hermes/.
Is Hermes Agent safe on a production server? The defaults already use smart approvals, a hardline blocklist, and a gateway that denies every user without an allowlist. For a server with client data, the documentation recommends manual approvals, a deny list, and the docker or ssh terminal backend.
Which platforms are supported? Tier 1: macOS Apple Silicon, Windows 10 and 11, Linux and WSL2, and Docker containers. Tier 2: Android Termux and Nix. Not supported: Intel macOS and installs through pip, brew, and AUR.
Next step
The 4-step test above gives you a cost figure and 5 real outputs in 1 week. The limit that stays: answer quality follows the model, and security follows the person who owns the approvals. When your team has passed the checklist and wants an install on a VPS with the provider, gateway, memory, skills, and SOP, the Hermes Agent Installation Service delivers it in 2 to 5 working days. When you are still weighing the scope, book a 60-minute Technical Scoping Session and bring 1 repeated task.
Sources
- Hermes Agent: official site
- Hermes Agent: Installation
- Hermes Agent: Platform Support
- Hermes Agent: Features Overview
- Hermes Agent: Hermes Desktop
- Hermes Agent: Messaging Gateway
- Hermes Agent: Running Hermes on a Personal or Work Machine
- Hermes Agent: FAQ
- Hermes Agent: Nous Portal
- Nous Portal: Manage Subscription
- Nous Portal: Privacy Policy
- GitHub: NousResearch/hermes-agent
- OpenClaw: What is OpenClaw?
