Hermes Agent: Pick it and plan

This is page 1 of 8 in the Hermes Agent guide, for a reader who picked Hermes Agent instead of Agent Zero. It covers who Hermes suits and what to sort out before you install it. It also covers which models Hermes works with, what they cost, and the catches.

Hermes Agent is made by Nous Research, and one install gives you a desktop app, a terminal program, and a way to chat with it from your phone.

The big difference from the main guide is where commands run. Agent Zero works inside a container, but Hermes works on your real computer, so the commands it runs can reach your real files. That is why page 2 puts the safety switches before the first job.

Facts here were checked on 14 November 2026, and Hermes changes often.

Pick this if

• You have a Mac with an Apple chip (M1 or newer), or a Windows 11 PC on version 22H2 or later.
• You want one program that can later answer from your phone and run jobs on a schedule.
• You are willing to change a few settings in a text file before the first real job.

Do not pick this if you have an Intel Mac. Hermes has no desktop app for Intel Macs, and its install guide calls Intel Macs "not a supported platform," so use the main guide instead.

Do not pick this if you want a wall between the agent and your files. Agent Zero has that wall by default, and Hermes does not.

Before you start

The Hermes program is free, and its license is MIT, which means anyone may use it at no cost. The model is a separate bill, and Pick your model and budget explains models, keys, and prices.

Start with a free Google Gemini key from Google AI Studio. You do not need a card, but Google may use free prompts to improve its products.

Three things to know first.

• Local models need a big memory window. A model running on your own computer must handle at least 64,000 tokens of context, which is how much text the model can keep in mind at once. Hermes refuses to start with less, and Ollama often starts much lower, sometimes 4,096.
• Claude Pro will not work through the built-in Claude login. Hermes's docs say Claude Pro "cannot use this path," and they point Pro users to a pay-as-you-go API key instead.
• Claude Max spends only extra credits. The built-in Claude login works only on Claude Max with extra usage credits bought on top, because Hermes does not use the usage that comes with Max. Every Hermes reply bills as extra usage, so you can get a separate bill even though you already pay for Claude.

Use your everyday Windows or Mac account, and do not run Hermes as an administrator, because Hermes runs commands as whatever account starts it.

Models it works with, and what they cost

Hermes needs at least one provider before it will answer, and it works with many of them. The full price table is on Pick your model and budget, so this section only covers what is different about Hermes.

• Google Gemini, with a key. This guide's starting point is a free Gemini key, which Hermes reads as GEMINI_API_KEY. Hermes's provider page warns that free-tier keys can run out after a handful of jobs, because Hermes may call the model several times for each message. A paid Google AI Pro or Ultra plan cannot sign in, so Gemini always needs a key.
• Nous Portal. Nous Research sells its own model plan, which also covers web search, image tools, and a cloud browser through one sign-in. Its free tier is $0 and only has free models. Plus is $20 a month, Super is $100 a month, and Ultra is $200 a month. Every paid plan includes one Hermes Cloud agent.
• OpenRouter. Hermes's setup table lists OpenRouter as a default choice, and it reaches hundreds of models through one key that you top up.
• Anthropic Claude. A pay-as-you-go API key works and bills per token, separately from any Claude plan. The built-in Claude login needs Claude Max plus bought extra credits, as described above.
• A local model. Ollama, LM Studio, or llama.cpp can run a model on your own computer for no per-use cost. The 64,000-token window and the setup make this a project rather than a first step, and page 7 covers it.

The Hermes FAQ also lists OpenAI, z.ai, Kimi, and MiniMax, and any service that speaks the OpenAI-compatible API.

The catches

The research scored Hermes 68.0 out of 100. That is third place, 2.0 points behind Agent Zero (70.0) and half a point behind Atomic Agent (68.5). See the scores.

It lost on safety, scoring 6 out of 20, against Agent Zero's 14. Commands run on your real computer, commands not on the dangerous list run without asking, and "smart" approvals let a second model say yes for you. It scored high for what it can do, 27 out of 30. One install can chat, run commands, edit files, browse, remember, write its own skills, run schedules, and reach many chat apps.

Other catches:

• The approvals have limits. When you click yes, the command runs with all of your account's power, and manual approval asks only about commands on the dangerous list.
• Docker turns off the questions. Hermes can run commands in a container instead, like Agent Zero. If you switch to that, Hermes stops asking, because the container is the wall.
• Its pages disagree. One Hermes page still says Windows users must use WSL2, but the install guide says Windows works without it, so follow the install guide. The oldest working macOS is listed as both 12 and 13.
• The installer adds screen control. The command install adds a browser and a tool that can drive your real screen and mouse. To skip the screen-control tool, run the Windows command as & ([scriptblock]::Create((irm https://hermes-agent.nousresearch.com/install.ps1))) -SkipComputerUse, or add --skip-computer-use after bash -s -- on a Mac.
• Memory can be wrong. Hermes keeps two small notes files about you and your setup. Its docs say weaker models, including small local ones, often claim they saved something when they did not.
• No Gemini plan sign-in. A paid Google AI Pro or Ultra plan cannot sign in to Hermes, so Gemini needs a key.
• Local models are a project. You need the 64,000-token window, the right server settings, and a big computer.
• Updates come fast. A normal install gets the stable version, 0.23.1 on the day checked, not the daily test builds.

Sources

Research in this repo:

• Hermes raw notes
• Agent landscape and scores
• Pick your model and budget

Official Hermes pages (checked 14 November 2026):

• Homepage and downloads: https://hermes-agent.nousresearch.com/
• Installation guide (opened again for this page on 14 November 2026; the --skip-computer-use and -SkipComputerUse switches): https://hermes-agent.nousresearch.com/docs/getting-started/installation
• Windows (native) guide (passing switches to the install script with [scriptblock]::Create): https://hermes-agent.nousresearch.com/docs/user-guide/windows-native
• Platform support: https://hermes-agent.nousresearch.com/docs/getting-started/platform-support
• Providers (Gemini, Claude, local models): https://hermes-agent.nousresearch.com/docs/integrations/providers
• Nous Portal: https://hermes-agent.nousresearch.com/docs/integrations/nous-portal
• FAQ (list of providers): https://hermes-agent.nousresearch.com/docs/reference/faq
• Memory: https://hermes-agent.nousresearch.com/docs/user-guide/features/memory
• Source code and README: https://github.com/NousResearch/hermes-agent

Next

Continue to Install and first job.