Is Hermes Agent free? What it really costs, free models and Nous Portal

The software is free; here is where money goes

Hermes Agent is made by Nous Research and published under the MIT licence. That licence lets anyone use, copy, modify and share it, commercially or not, without paying a fee, provided the copyright and permission notices are included in copies or substantial portions of the software. You do not need to buy a licence to run it.

A running agent still has three possible bills:

1. The AI model. Hermes is the agent, not the model: it sends each request to a model you choose. Its cost depends on the model, its pricing and how much you use it. One message from you can turn into several model calls, because the agent calls tools, reads the results and carries on. Background tasks such as generating titles, summarising conversations and analysing images can also make model calls. Many default to your main model, but image analysis can use a separate vision model or another configured provider, so check the auxiliary model settings when estimating costs. Scheduled jobs use the model as well, even while you are not chatting.
2. The machine. Your own computer costs nothing extra, but the agent only answers while it is switched on and online. Keeping it reachable around the clock means a server or a hosted agent; our guide to cheap Hermes Agent hosting covers those costs.
3. Optional tools. Web search, image generation, voice, cloud browsers and external memory can each be connected to a paid service. Several have free routes, covered below.

Paying for the model: the three usual routes

• A pay-per-token API key. You put a provider’s key in ~/.hermes/.env or enter it in hermes model, and the provider bills you for every token sent and received. Hermes supports a long list of providers this way. Prices differ a lot between models; see our pages on using Claude and using DeepSeek for two common ends of that range.
• A subscription you already pay for. Some providers let you sign in with a consumer plan instead of a key. For example, hermes model offers “ChatGPT or Codex Subscription”, which signs in with your ChatGPT account. The Hermes docs are open about the limits here: how that usage counts against a ChatGPT plan is not documented, and each provider has its own rules, so check the provider’s own billing page before relying on it.
• Nous Portal. A single Nous Research account that covers many models and some tools. It has a free plan and paid plans; the section on Nous Portal below explains both.

Two built-in safeguards help. Type /usage in a chat to see a breakdown of tokens and cost for the session. And when you pick a model whose known price is above US$20 per million input tokens or US$100 per million output tokens, Hermes shows an expensive-model warning and asks you to confirm.

Free ways to run the model

These are the routes the 0.22.5 source and docs support that can cost nothing for the model itself:

• Local models on your own hardware. Local model inference runs on your computer without a per-token provider bill. Remote tools and separately configured cloud models can still send data off the machine. The Hermes Desktop app has a built-in Local Models section that downloads llama.cpp and a model sized for your machine. You can also point Hermes at Ollama, LM Studio, llama.cpp’s server or vLLM through hermes model; our guide to Hermes Agent with Ollama walks through that setup.
• The Nous Portal free plan. Signing up for a Nous account is free. On the free plan, Hermes’s model picker only lets you choose models the Portal prices at zero; paid models are shown greyed out. When we read the Portal’s homepage on 30 October 2026, a handful of models were listed at $0.00 per million tokens. Which ones are free changes over time, so check the list in hermes model.
• The Desktop app’s first-run free tier. On a first install of Hermes Desktop, before any model is set up, Hermes can start you on a single free Nous model without signing in; it appears as “Nous · free tier”. It is a way to try Hermes, not a long-term plan: Nous can pause or limit it, and it stops applying as soon as you choose a model of your own.
• Providers with a free tier. Two providers Hermes supports advertised free use on their own pages when we read them on 30 October 2026. Google’s Gemini API pricing page lists a free tier with free input and output tokens on some models, and says content sent on the free tier is used to improve Google’s products; in Hermes this is the gemini provider with GOOGLE_API_KEY or GEMINI_API_KEY. NVIDIA’s build.nvidia.com offers “free inference with leading models”; in Hermes this is the nvidia provider with NVIDIA_API_KEY.

Two routes that look free are not, at the time of writing. The Hermes docs mention a small monthly Hugging Face credit, but Hugging Face’s own pricing page (read 30 October 2026) lists no monthly credits for free accounts. And Ollama Cloud, unlike Ollama on your own machine, charges per token for its cloud models.

The catch with free and local models

An agent asks more of a model than a chat window does. It has to choose tools, fill in their inputs correctly and keep a plan going over many steps. That is where free and local models most often fall short:

• Context window. Hermes normally refuses to start a session on a model with less than 64,000 tokens of context, because the instructions, tool descriptions and conversation need the room. Local servers often default to less, so you may need to raise the setting before anything works.
• Hardware. The Hermes docs say a graphics card with 8 GB or more of memory runs the small models in the Desktop catalogue comfortably, and 16 GB or more runs 27 to 35 billion parameter models at high quality. A model that spills into ordinary system memory still works, but slower.
• Tool use and speed. Smaller models are more likely to call the wrong tool, give up half way, or take a long time over a multi-step job. A model that is fine for chatting can be poor at agent work. Even Nous’s own Hermes 4 chat models are not recommended inside Hermes Agent, according to the Nous Portal docs, because they are tuned for chat and reasoning rather than rapid tool calling.
• Limits on free tiers. Free hosted tiers are rate-limited. The Hermes docs warn that a free Gemini key can run out after a handful of agent turns, and recommend a billing-enabled project for real use.
• Your data. A local model keeps its inference on your machine, but remote tools and auxiliary cloud models can still receive data. For hosted services, read the provider’s terms, including Google’s note above.

A sensible pattern is to try Hermes on a free option, then move the jobs you care about to a capable paid model once you know what you use it for.

Optional tools: which need paid keys and which are free

• Web search. Works on a fresh install with nothing set up: Hermes rotates across the public free tiers of Exa, Parallel, Firecrawl and Keenable, and moves to the next when one is busy. SearXNG (if you run it yourself) and DuckDuckGo are also free. Paid keys from those vendors, or from Perplexity, give you steadier service. web.keyless_fallback: false turns the free rotation off.
• Text-to-speech. The default, Edge TTS, is free and needs no setup. NeuTTS, KittenTTS and Piper run on your own machine for free. Cloud voices such as ElevenLabs and OpenAI need a key or, for OpenAI, a Nous Portal account with tool access.
• Voice notes to text. Local faster-whisper transcription runs on your machine at no cost; cloud transcription needs a provider key or a Nous Portal account with tool access.
• Image generation. It normally requires a funded provider key or a paid Nous Portal subscription. Some Nous accounts receive a small free managed-tool allowance that covers supported image generation; Hermes offers setup on first use when that allowance is available.
• Browser. Driving a local Chrome or Chromium on your own machine is free. Cloud browsers such as Browser Use and Browserbase need an account and key.
• Memory. The built-in memory files are enabled by default and have no separate service fee; you can disable them in configuration. External memory providers are optional; several, such as Mem0 and Honcho, are free if you host them yourself and paid as a cloud service.

What Nous Portal is, and what it costs

Nous Portal is Nous Research’s account and subscription service, and the Hermes docs call it the recommended way to run Hermes Agent. One sign-in gives Hermes access to a catalogue of several hundred models from many labs, billed to your Nous account instead of a separate account with each lab. Paid plans also unlock the Tool Gateway, which routes web search, image generation, text-to-speech and cloud browsing through Nous, so you do not need an account with each of those services.

On a fresh install, one command signs you in, lets you pick a model and offers to turn on the Tool Gateway tools your plan includes:

hermes setup --portal

On an existing install, run hermes model and pick “Nous Portal”; your other providers stay configured. hermes portal runs the same sign-in on its own, and hermes portal info shows whether you are signed in and which tools go through Nous. The sign-in opens a browser, and Hermes keeps the resulting credential in ~/.hermes/auth.json. If you use several Hermes profiles, one sign-in covers all of them.

These are the plans listed on portal.nousresearch.com when we read it on 30 October 2026:

The same page says large models have their own monthly limit on paid plans, that calls made with a Portal API key always use pay-as-you-go credits rather than the included usage, and that credits can be topped up. It also lets you connect an eligible ChatGPT account for inference without a paid Nous plan, with tools and cloud hosting charged separately. Plans change, so check the Portal before you subscribe; hermes portal open opens the subscription page.

Hermes Agent costs on ReadyHermes

ReadyHermes runs your own Hermes Agent on our servers, always on, with the AI model already set up, so there is nothing to choose or connect before your first message. Here is how the costs above work with us:

• A free trial of one hour. Sign up and your agent is ready in about 10 seconds. The hour starts when it first runs, and the trial uses AI we include, so you can try real jobs straight away. When it ends, your agent pauses: choose a plan within 24 hours to keep it, along with its memory and connected apps, or it is deleted.
• AI included in every plan. Plans come with a monthly amount of included AI, and you can switch between the included models in the workspace; larger models use the allowance up faster.
• Or your own provider key. During the trial or on a paid plan you can switch to your own key from OpenRouter, OpenAI, Anthropic, Google Gemini, DeepSeek, Groq, Qwen, Z.AI, Moonshot or MiniMax, or another supported OpenAI-compatible service. We send one small test message first, the provider bills you directly, and we put no AI limit on your own key. “Use included AI” switches back. Switching restarts the assistant, which takes about a minute.

Two things on this page ReadyHermes does not offer: it does not run models on your own computer, and Nous Portal is not one of the providers in the key panel. For plan prices and what each plan includes, see cheap Hermes Agent hosting.