Nous Research
Nous Research is an American open-source AI lab. It trains and releases the Hermes model family, ships Hermes Agent, a self-improving autonomous agent under MIT licence, and runs Nous Portal, one subscription covering 300+ frontier models.
What is Nous Research?
Nous Research is an American open-source AI laboratory that presents itself as a leader in the American open source AI movement. Its stated mission is to advance human rights and freedoms by creating and proliferating open source language models, and its declared research focus covers model architecture, data synthesis, fine-tuning and reasoning.
The lab works on three fronts. First, it trains and publishes open-weight models: the Hermes family (Hermes 4 in 405B, 70B and 14B variants, and Hermes-4.3-Seed-36B, post-trained entirely on the Psyche network), the Nomos 1 mathematics model, NousCoder-14B, DeepHermes-3, YaRN, Genstruct-7B and Obsidian, alongside public datasets and technical reports. The Releases page lists 43 dated outputs going back to July 2023.
Second, it ships Hermes Agent, an autonomous agent released under the MIT licence, with native applications for macOS 12+, Windows 10/11 and Linux. Nous positions it as neither a coding copilot tethered to an IDE nor a chatbot wrapper around a single API, but as an agent that becomes more capable the longer it runs. It carries a closed learning loop: agent-curated memory, autonomous skill creation, skill self-improvement during use, cross-session recall and dialectic user modelling through Honcho. It runs on six terminal backends — local, Docker, SSH, Daytona, Singularity and Modal — two of which hibernate when idle, and it reaches the user across more than twenty messaging surfaces, from Telegram and Discord to Slack, WhatsApp, Signal and Microsoft Teams, all sharing one memory. It exposes over sixty built-in tools, supports the Model Context Protocol and can spawn isolated subagents with their own terminals.
Third, Nous Portal is a subscription gateway where one OAuth login replaces separate accounts and API keys across labs. It opens a catalogue of 300+ agentic models from Anthropic, OpenAI, Google, DeepSeek, Qwen, xAI and others, plus a Tool Gateway bundling web search and extraction, image generation, text-to-speech, cloud browser automation and an optional cloud terminal sandbox.
Alongside these, Psyche is an open infrastructure that decentralises training across underused hardware, built on the DisTrO and DeMo optimisation work and coordinated on the Solana blockchain.
What it does
- Run a persistent autonomous agent on a server, a VPS or serverless infrastructure
- Drive that agent from Telegram, Discord, Slack, WhatsApp, Signal or twenty other messaging surfaces
- Reach 300+ frontier models from Anthropic, OpenAI, Google and others through a single subscription
- Schedule recurring work in plain language: reports, backups, briefings, run unattended
- Download and run open-weight Hermes, Nomos and NousCoder models locally
- Search, extract and browse the web, generate images and speak through one built-in tool gateway
- Contribute GPU capacity to distributed model training through the Psyche network
When to use Nous Research / When not to
A quick filter to help you decide if Nous Research is the right fit.
When to use Nous Research
- AI and machine-learning researchers who need open-weight models they can download, inspect and post-train themselves
- Developers looking for a self-hostable autonomous agent they can script, sandbox and run on their own servers
- Engineering teams that want one billing relationship instead of separate accounts with every model lab and tool provider
- Technically comfortable users at ease with a command line, Docker, SSH and configuration files
- Academic labs and students working on distributed training, reinforcement-learning environments or reasoning benchmarks
When not to use Nous Research
- Organisations that require a documented GDPR framework, a data processing agreement or a published subprocessor list
- Buyers who need to know what a subscription costs before creating an account, since no price is published
- Non-technical users expecting a ready-to-use consumer interface with no installation step
- Anyone whose conversations must never be used to train or fine-tune models, as the privacy policy claims that use with no documented opt-out
- People under 18, since the terms of service require users to be at least 18 or to hold their legal representative's authorisation
How to use Nous Research
A typical end-to-end flow, from setup to results.
- Decide which surface you need: the open-source Hermes Agent, a Nous Portal subscription for model access, or simply the open-weight models on Hugging Face
- For the desktop agent, download the macOS installer or the Windows installer from the Hermes Agent site
- For a command-line only install on Linux, macOS, WSL2 or Android via Termux, run the install.sh one-liner published on the site
- On native Windows without the desktop app, run the equivalent PowerShell install command
- Run 'hermes setup --portal': a single OAuth covers one model plus all four Tool Gateway tools (web search, image generation, text-to-speech, browser)
- If you have no subscription yet, sign up on the Nous Portal plans page first, then come back and run the setup command
- Start working with 'hermes chat', and switch model mid-session with the '/model' command
- To add the Portal alongside providers you already use, run 'hermes model' and pick Nous Portal from the provider list; existing providers stay configured
- Configure the messaging gateway to reach the agent from Telegram, Discord, Slack, WhatsApp or another supported platform
- For a headless or SSH host, follow the documented OAuth-over-SSH pattern with ssh -L port forwarding, since the loopback callback runs on the machine hosting the agent
Pros & Cons
Pros
- Genuinely open: the agent is MIT-licensed, and models, datasets and technical reports are published on GitHub and Hugging Face
- One subscription replaces five separate third-party accounts for search, images, speech, browser automation and sandboxing
- A permanent free tier exists alongside the paid Plus, Super and Ultra tiers
- Extensive documentation, available in English and Simplified Chinese, with machine-readable llms.txt files for ingestion
- The agent is not tied to a laptop: six execution backends, two of which cost almost nothing while idle
- An identified, funded company with a visible research output rather than an anonymous product page
- Unusually candid product guidance: the documentation explicitly advises against using Nous's own Hermes 4 models for agentic work
Cons
- No price is published anywhere reachable: the plans page sits behind an anti-bot challenge, so the cost is unknown before signup
- The privacy policy states that conversations are used to train and fine-tune models, with no documented opt-out
- That policy dates from November 2024 and is largely written around WorldSim, well behind the current product line
- No GDPR mention at all, no data processing agreement, no Article 27 representative and no enumerated user rights
- No postal address, no contact page and no legal notice page anywhere on the site
- The terms of service and privacy policy live on a subdomain that will not open without a browser
- Installation via a piped curl script and configuration by YAML file assume a fair degree of technical comfort
Pricing & Plans
A permanent free plan exists: Nous Portal is offered on four tiers named Free, Plus, Super and Ultra, and all paid tiers are stated to include monthly credits for use in Hermes Agent together with access to 300+ models and built-in tool use. Nous Research does not publish the amount of any paid tier on a publicly reachable page, so no entry price and no currency can be quoted here. Hermes Agent itself is free and open source under the MIT licence, and it can be run against third-party providers without any Nous subscription.
- Free — the no-cost Nous Portal tier
- Plus — paid tier including monthly credits for Hermes Agent
- Super — higher paid tier
- Ultra — highest paid tier
- No amount is published for any paid tier on a publicly reachable page
Data, GDPR & hosting
A consolidated view of how Nous Research handles your data.
GDPR overview
There is no mention of the GDPR anywhere: not in the privacy policy, not in the terms of service, and not in the product documentation. The site names no legal basis for processing, enumerates none of the data subject rights (access, rectification, erasure, portability, objection), designates no Article 27 representative and no data protection officer, and offers no data processing agreement. The privacy policy is a short document effective 13 November 2024 that predates much of the current product line and is largely written around WorldSim. The terms of service place exclusive jurisdiction with the United States Federal Courts in San Francisco, California. This is silence rather than a refusal, but silence is what a European user has to work with, and it means no GDPR commitment can be relied upon here.
Who owns the data?
The published documents say little about ownership. The privacy policy, effective 13 November 2024, states that Nous Research collects conversation texts, usage details and login information, and claims to process them in an anonymised fashion. It states that Nous Research does not disclose personal information recorded on its systems to outside parties, while reserving the right to share anonymised session data such as average session lengths with third parties. The terms of service identify the counterparty as Nous Research, Inc., a Delaware corporation. No clause in the collected material sets out who owns customer content, and no licence grant over user data is spelled out either way.
Reuse rights
The privacy policy lists four uses for the information collected: delivering the web application, training and fine-tuning models, research and development for WorldSim and other Nous Research initiatives, and analytics. Training on user conversations is therefore an explicitly claimed use, and no opt-out is documented anywhere on the site, in the product documentation or in the legal pages. The policy also states that personal information is used only for authentication purposes, a statement that sits awkwardly beside the training clause. On the user's side, anything typed into a third-party model routed through the platform falls under that provider's own terms and privacy policy, which Nous Research explicitly disclaims responsibility for. Nothing in the collected documents restricts what a user may do with the outputs they generate, and the agent's own code is released under the permissive MIT licence.
Data retention & training
Hosting summary
No hosting country and no hosting region is declared in any published document. The privacy policy limits itself to saying that all information is stored in compliance with industry security standards, which says nothing about jurisdiction. What can be observed from outside is only infrastructure: the domain resolves to an address geolocated in the United States on Amazon's network, served through an anycast node, and the portal subdomain runs on Vercel. That is hosting of the website, not a statement about where user data lives. On the agent side the picture is different and more favourable: Hermes Agent executes locally or on infrastructure the user chooses themselves, among local, Docker, SSH, Daytona, Singularity and Modal backends, so the memory and working files stay wherever the user put them. Terms of service place exclusive jurisdiction with the United States Federal Courts in San Francisco, California.
Things to keep in mind
Risks and trade-offs to weigh before adopting Nous Research.
- The privacy policy states that your conversations are used to train and fine-tune models, and no opt-out is documented, so treat anything you type as potential training material
- That policy is dated 13 November 2024 and is largely written around an earlier product, so it may not describe how the current agent and portal actually handle data
- No GDPR mention, no data processing agreement, no Article 27 representative and no enumerated user rights leave a European user with no documented recourse
- The terms of service and privacy policy sit on a subdomain that blocks automated access, which makes the legal terms harder to review than they should be
- No postal address, contact page or legal notice is published, so the only channel is a support email found inside the privacy policy
- Installing through a piped curl script executes remote code on your machine, which deserves a deliberate decision rather than a reflex in a managed environment
- An agent with persistent memory, scheduled automation and its own terminal is genuinely powerful and correspondingly easy to over-trust: review what it is allowed to run before letting it work unattended
Setup & Integrations
Technical difficulty
Moderate, and it depends on how far you go. Native installers for macOS and Windows make the desktop agent straightforward, and a single command, 'hermes setup --portal', configures a model and the four gateway tools in one pass. Beyond that the tool expects some comfort: a command-line install through a piped script, a config.yaml file to edit, a choice of terminal backend, messaging gateways to connect and optional MCP servers to wire up. Headless or SSH setups are documented but need port forwarding. Non-technical users will struggle; a developer will not.
Deployment
Integrations
Supported languages
Behind Nous Research
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
Is Hermes Agent free?
How much does a Nous Portal subscription cost?
Are my conversations used to train models?
Is Nous Research GDPR compliant?
Which systems does Hermes Agent run on?
Can I use the agent without a Nous subscription?
Is there a minimum age?
Where is my data hosted?
Is there an API?
Which models does Nous Research publish?
Should you pick Nous Research?
Nous Research is one of the more substantial entries in the open-source AI landscape, and the evidence is easy to check: an identified company, Nous Research, Inc., incorporated in Delaware; forty-three dated releases; model weights and datasets published on Hugging Face; source code on GitHub under the MIT licence; and technical reports hosted on its own site. The three product surfaces hold together coherently. The lab trains open models, Hermes Agent turns them into an autonomous assistant that runs wherever the user puts it, and Nous Portal collapses a dozen billing relationships into one. For a researcher, an infrastructure engineer or a developer who wants weights rather than a hosted black box, the offer is unusually complete.
The reservations are all on the paperwork rather than the product. The privacy policy dates from November 2024, is still largely written around an earlier product, and states plainly that conversations are used to train and fine-tune models without documenting any way to opt out. The GDPR is never mentioned, no data processing agreement exists, and no postal address, contact page or legal notice appears anywhere. The legal pages themselves are only reachable through a subdomain that refuses automated access. Pricing is equally opaque: four tiers are named, no amount is published.
The result is a tool that technical and research-minded users will find genuinely valuable, and that a compliance officer will find hard to sign off. Anyone bound by European data protection requirements, or who simply needs to know a price before committing, should treat the documentation gap as the real obstacle rather than the technology.
- Choosing a selection results in a full page refresh.
- Opens in a new window.