Hermes Agent

What is Hermes Agent?

Hermes Agent is an open-source AI agent from Nous Research that you install on your own computer, a server, or a cloud instance and then reach through a terminal, a desktop app, or messaging apps. Nous Research describes it as an open-source autonomous agent that learns across sessions and writes its own skills from experience.

The name overlaps with other Nous Research work: the company has also released language models of its own. This page covers the agent software; models, from Nous or anyone else, are something it connects to.

Hermes Agent is free and open source under the MIT license, while model providers and optional hosted services have their own pricing. It can run on your own hardware, on a remote machine over SSH, inside containers, or as a paid hosted instance managed by Nous.

The project is widely followed: its GitHub repository showed 249,322 stars as captured on September 27, 2026. In July 2026, technology press reported that Nous Research, the startup behind the open-source Hermes agent, was finalizing a funding round at a $1.5 billion valuation, citing three sources with knowledge of the deal.

Core features

Memory and skills

• Bounded memory files: Hermes Agent has bounded, curated memory that persists across sessions, kept in two files — MEMORY.md for the agent's own notes (2,200 characters, about 800 tokens) and USER.md for your profile and preferences (1,375 characters, about 500 tokens).
• Session search: all CLI and messaging sessions are stored in a local SQLite database under the Hermes home folder with FTS5 full-text search, so the agent can look up past conversations.
• Self-written skills: the agent can create, update, and delete its own skills through a skill-management tool, which the documentation calls its procedural memory.
• Skill format: skills are on-demand knowledge documents loaded only when needed and compatible with the agentskills.io open standard.

The memory limit is deliberate. Memory does not auto-compact; when a write would exceed the limit, the memory tool returns an error instead of silently dropping entries, and the agent has to consolidate or remove entries itself. Skills carry the longer procedures, which is why the project pairs small memory files with a growing skill library rather than one large memory store. The README describes this as a closed learning loop, with autonomous skill creation after complex tasks and skills that self-improve during use.

Tools and delegation

• Built-in tools: web search, browser automation, vision, image generation, text-to-speech, and multi-model reasoning.
• Subagents: isolated subagents with their own conversations, terminals, and Python RPC scripts for zero-context-cost pipelines.
• MCP servers: the documentation includes a guide to connect any MCP server for extended capabilities.
• Model switching: the README says you switch with hermes model — no code changes, no lock-in.

Messaging gateway and scheduling

The gateway is a single background process that connects to all your configured platforms, handles sessions, runs cron jobs, and delivers voice messages. Scheduling runs through it: you can schedule tasks to run automatically with natural language or cron expressions, and results can be delivered back to a chat. Because both scheduled work and chat replies depend on that process, the gateway is what turns Hermes Agent from a terminal tool into an always-on assistant.

Execution backends

Where commands run is configurable. The homepage lists five backends — local, Docker, SSH, Singularity, and Modal — with container hardening and namespace isolation, while the README lists seven terminal backends, adding Daytona and Vercel Sandbox. The two official pages differ on the count, so the exact set depends on the version you install.

Guide

The documented path is short, but the project stresses getting one working chat before adding more.

1. Install: download Hermes Desktop for macOS or Windows, or use the one-line terminal installer on Linux, macOS, or WSL2. Native Windows runs Hermes without WSL, with the CLI, gateway, TUI, and tools working natively.
2. Pick a setup mode: a fresh hermes setup offers a Quick Setup that logs into Nous Portal, a Full Setup where you walk through every provider, tool, and option yourself (bring your own keys), and a minimal Blank Slate mode.
3. Connect a model: you need at least one provider configured to use Hermes. For a local model, the quickstart says to verify the endpoint, model name, and context length.
4. Verify a plain chat first: the documentation's rule of thumb is that if Hermes cannot complete a normal chat, do not add more features yet.
5. Add the gateway: run the gateway setup after the CLI works to connect Telegram, Discord, Slack, or another platform.

When moving to another computer, the homepage says to keep a backup of your Hermes data directory. Users coming from OpenClaw can run a migration command that imports settings, memories, skills, and API keys.

Use cases and real-world examples

The official documentation hosts a curated gallery of community user stories, linked to public posts, issues, videos, and gists. Its category counts show where reported use concentrates: of 326 stories, 77 are dev workflow, 49 personal assistant, 32 integrations, 26 creative, and 23 business operations.

• Scheduled briefings and maintenance: the homepage pitches natural-language scheduling for reports, backups, and briefings, running unattended through the gateway.
• Remote work from chat: the README frames it as not tied to your laptop — you can talk to it from Telegram while it works on a cloud VM.
• Solo software projects: one story in the gallery describes a developer whose agent files its own tasks, writes code, runs QA, deploys, and keeps documentation current while the developer reviews over Telegram.
• Parallel workstreams: delegation to isolated subagents lets multi-step pipelines run without filling the main conversation's context.

These are vendor-curated examples. They show what users say they built, not measured outcomes.

Who is it for

Hermes Agent fits people who are comfortable running a self-hosted AI agent themselves and want an assistant that keeps state over weeks rather than a single chat window.

• Developers and tinkerers who want a terminal agent that can run commands, edit files, and use their own model keys. The security documentation lists the local backend as best for development and trusted users.
• People who want an always-on bot: the quickstart is written for anyone setting up Hermes for a team, bot, or always-on workflow.
• Privacy-minded users with local hardware: the provider guide describes Ollama as best for quick local experimentation, privacy-sensitive work, and offline use.
• Research teams: the README lists batch trajectory generation and trajectory compression for training tool-calling models.

It is a weaker fit if you want a zero-setup consumer app. The base desktop app needs no account, but meaningful use still requires connecting a model provider, and a model with too small a context window is rejected at startup.

Platforms

Mac processor support is described differently across official pages. The installation guide says the macOS installer is Apple Silicon only and that Intel Macs are not a supported platform. The platform support page says the Hermes-Setup.dmg bootstrap installer is Apple Silicon only, but that Intel Macs use the darwin-x64 desktop bundle instead, or can install the CLI and run the desktop app from it.

On Android, the package is a prerelease build with Python, Node, and the TUI; you run the gateway in a Termux session, and Android can terminate background processes.

Messaging coverage is broad. The messaging guide lists Telegram, Discord, Slack, WhatsApp, Signal, SMS, Email, Home Assistant, Mattermost, Matrix, DingTalk, Feishu/Lark, WeCom, Weixin, BlueBubbles (iMessage), QQ, Yuanbao, Microsoft Teams, LINE, ntfy, and the browser.

Pricing

The agent itself costs nothing to download or use; spending comes from models, hosted tools, and optional hosting. You can bring your own provider keys, or pay Nous through Nous Portal. No account is needed to run the base desktop app; signing in to Nous Portal adds always-on cloud agents, free and discounted models, and bundled tool usage credits.

Nous Portal plans

The homepage says all paid tiers include monthly credits for use in Hermes Agent and access to 200+ cutting-edge models, while the quickstart says one subscription covers 300+ models plus the Tool Gateway. The two official pages give different model counts. The Tool Gateway is included with every paid Nous Portal subscription and routes web search, image generation, text-to-speech, and cloud browser automation; all four are pay-as-you-use billed against your Nous subscription.

Hermes Cloud hosting

Hermes Cloud is the optional hosted version, billed daily from Nous credit based on usage:

Inference and tool usage are billed separately, on top of the instance price. A stopped instance is charged storage only. The Nous Portal terms state that all fees are non-refundable, with no refunds or credits for unused or partially used periods.

Hermes Agent alternatives

The most direct comparison is OpenClaw. Technology press described Hermes as a competitor Nous Research released weeks after OpenClaw's agent went viral, noting that Hermes shipped with built-in skills such as web search, coding, and image understanding.

• OpenClaw: an open-source AI assistant that runs on your own computer and works through Discord, iMessage, Slack, Teams, Telegram, WhatsApp, and more than 20 other channels, with native apps for macOS, iOS, Android, Windows, and Linux. For anyone switching, Hermes can automatically import settings, memories, skills, and API keys from OpenClaw.
• Agent Zero: an open agent framework built around a Dockerized Linux desktop, a browser with DOM annotation, live document cowork, projects, skills, and plugins, with a bridge back to the host machine.

As an OpenClaw alternative, Hermes Agent is written mainly in Python and centers on memory plus self-written skills, while Agent Zero centers the work on a containerized desktop.

Limitations

Setup and model requirements

• Minimum model context: Hermes Agent requires a model with at least 64,000 tokens of context; models with smaller windows are rejected at startup. Local models with small default context settings need reconfiguring.
• Provider needed: the software is free, but it does nothing useful until at least one model provider is configured.
• Agent must be running: tasks that need an active agent will not run while a local agent is stopped.
• One agent per home folder: the memory documentation warns not to point two agent processes at the same Hermes home directory.

Execution safety

• Local backend has no isolation: in the security documentation's backend table, the local backend runs directly on the host, with dangerous-command checks as the main safeguard.
• Approval checks can be disabled: before executing any command, Hermes checks it against a curated list of dangerous patterns and asks for approval on a match, but turning approvals off disables all safety prompts, which the documentation reserves for trusted environments.
• Skills are written freely by default: by default the agent writes skills freely, including from the background self-improvement review; a write-approval setting can require human review first.
• Gateway access is closed by default: with no user allowlists configured, the gateway logs a startup warning and denies all unauthorized users.

Windows security software

The README warns that antivirus tools such as Bitdefender and Windows Defender may quarantine the bundled uv.exe and calls this a false positive, with whitelisting steps. Public issue reports from separate users go further: one describes Microsoft Defender quarantining the regenerated Windows CLI launcher, and another describes Smart App Control blocking unsigned Hermes executables outright, with no documentation guidance at the time of the report.

Privacy and rights

• Local storage: in a self-hosted install, memory files are stored under the Hermes home folder (~/.hermes/memories/), next to the session database.
• Nous services: the Nous Portal privacy policy covers Hermes and other agents and tools made available by Nous. It says prompts, inputs, and outputs may be shared with third-party model and infrastructure providers as needed to provide the services, and that personal information may be disclosed to providers supporting development, evaluation, and training of its AI technologies.
• Output ownership: under the Nous Portal terms, you own AI outputs generated specifically for you, subject to compliance and payment, while Nous retains rights in the underlying models.

FAQ

Q1. Is Hermes Agent free?

Yes. Costs come from the model provider you connect, optional Nous Portal plans, and optional Hermes Cloud hosting.

Q2. Is Hermes Agent the same as the Hermes language models?

No. Hermes Agent is agent software. It connects to models from many providers, including Nous Portal, OpenRouter, Anthropic, and local servers such as Ollama.

Q3. Can Hermes Agent run on local models?

Yes, through self-hosted endpoints such as Ollama or vLLM, provided the model offers at least 64,000 tokens of context. Provider access and optional integrations can still require network access.

Q4. What happens to my data when I stop the agent?

Stopping a local agent ends its running process, but does not delete its saved conversations, memories, or skills.