AI agents (Hermes)

Hermes Agent by Nous Research: Features and Business Rollout

11 min read Updated: Czytaj po polsku
Hermes Agent by Nous Research: Features and Business Rollout

Hermes Agent is a free, open-source (MIT-licensed) AI agent from Nous Research that you install on your own machine or server. It remembers context across sessions, writes reusable skills from the work it has done, and answers you on Telegram, Slack, WhatsApp, Microsoft Teams and more. For a business it works well as an operations assistant running scheduled jobs — provided you set up isolation, access allowlists and command approval from day one.

What is Hermes Agent?

Hermes Agent is an autonomous AI agent: you give it a goal, and it chooses its own tools — terminal, browser, files, APIs — and works through the steps until the job is done. It is built by Nous Research, the lab behind the Hermes family of language models. The official documentation is explicit that this is neither a coding copilot tied to an IDE nor a chat wrapper around a single API: it can run on a $5 VPS, a GPU cluster or serverless infrastructure such as Daytona or Modal, while you talk to it from Telegram.

The code lives in the public NousResearch/hermes-agent repository — mostly Python, MIT licence. It moves fast: at the time of writing (18 November 2026) the latest stable release is v0.23.2, published on 16 November, with daily “canary” builds alongside stable versions (GitHub releases). In June 2026 Nous Research also shipped Hermes Desktop as a public preview for Windows, macOS and Linux (The Decoder).

Glossary: terms you will meet

  • Skill — a SKILL.md file describing when to use it, the procedure, pitfalls and how to verify the result. Every installed skill also becomes a slash command, e.g. /github-pr-workflow.
  • Gateway — one background process that connects the agent to messaging apps and runs the scheduler.
  • Terminal backend — where the agent executes commands: locally, in Docker, over SSH, or in a cloud sandbox (Daytona, Modal, Vercel Sandbox).
  • SOUL.md — the agent’s identity file, placed first in the system prompt. It sets tone, role and default behaviour.

What can Hermes Agent do?

Core capabilities from the features overview and what they mean for a team.

Area What you get Why it matters at work
Memory MEMORY.md (agent notes, 2,200-character limit) and USER.md (user profile, 1,375 characters), loaded at session start The agent knows your conventions without being reminded every time
Session search Conversation history in SQLite with FTS5 full-text search Recall the details of a decision made weeks ago
Skills Compatible with the agentskills.io standard; created and patched by the agent through skill_manage Company procedures written once, executed consistently
Gateway 20+ platforms including Telegram, Discord, Slack, WhatsApp, Signal, email, SMS, Microsoft Teams and Google Chat The agent lives where your team already talks
Cron Jobs defined in plain language or cron syntax, delivered to any platform Reports and checks that run without anyone pressing a button
Subagents delegate_task spawns child agents with isolated context; three run concurrently by default Parallel research and splitting large jobs
Code execution execute_code runs agent-written Python that calls Hermes tools in a single model turn Fewer model calls in repetitive pipelines
Tools and MCP 60+ built-in tools, browser automation, MCP servers over stdio or HTTP Connect your CRM, database or shared drive
Models Nous Portal, OpenRouter, OpenAI, any endpoint, local models (Ollama, vLLM, llama.cpp) No lock-in to one provider
API An OpenAI-compatible HTTP endpoint that works with Open WebUI, LibreChat and similar front ends Use Hermes as the engine behind other apps

On top of that you get checkpoints (the agent snapshots the working directory before editing files, and /rollback undoes the change), voice mode, and Bot Mode, where several specialist agents sit in a group chat and respond to @mentions.

How does the learning loop work?

Nous Research markets Hermes as the agent with a built-in learning loop. That does not mean the model is retrained — the weights never change. The agent learns by writing knowledge into files you can read, edit and keep in Git. Three mechanisms do the work.

  1. Curated memory. According to the memory docs, both files are injected into the system prompt as a frozen snapshot when a session starts. New entries hit the disk immediately but only show up in the prompt next session. When memory is full, the tool returns an error and the agent has to consolidate or delete entries. Every entry is scanned for injection and exfiltration patterns.
  2. Session search. session_search returns the actual stored messages, not model summaries. Memory holds what must always be in context; search handles the details you only need occasionally.
  3. Skills as procedural memory. After solving an awkward problem, the agent can save a new skill or patch an existing one. The skills documentation insists skills should capture “lessons, not logs”. The agent first sees only names and descriptions (about 3k tokens) and loads the full text when it actually needs it, so a hundred skills do not flood the context window.

One detail matters for businesses: by default, skill writes are not reviewed (skills.write_approval: false). Turn it on and changes are staged for you to inspect with /skills diff and accept with /skills approve. You can also freeze critical skills with hermes curator lock. The /learn command turns a folder, URL or long document into a skill with an index, which is the quickest way to load your internal procedures.

Hermes Agent vs n8n vs a chatbot

They rarely compete. We cover the distinction in our guide to AI agents for business; here is the short version.

Criterion Website chatbot n8n Hermes Agent
How it works Answers questions from a knowledge base Runs a workflow you designed in advance Plans its own steps and picks tools
Predictability High within a narrow scope Very high Medium — depends on the model and guardrails
Best for FAQs, lead capture Rule-based processes: invoices, leads, data sync Open-ended work: research, analysis, ops that need judgement
Memory across sessions Usually none Only what you build into a database Built in: memory, session search, skills

The pairing that works best in practice: n8n handles the deterministic steps (fetch data, update the CRM, send the invoice) and Hermes steps in where judgement is needed — summarising, triaging edge cases, drafting a reply for approval. Because Hermes can expose an OpenAI-compatible endpoint, n8n can call it like any other model; in the other direction, the agent can trigger n8n workflows through a webhook or an MCP server.

What can a business use Hermes Agent for?

Typical first deployments — each runs on a single server and needs access only to the data it uses.

  1. Morning briefing. A 7:30 cron job pulls sales and analytics data over MCP, compares it with last week and posts five bullet points to a Slack channel.
  2. Team assistant in Teams or Slack. Onboarding, complaints handling and publishing checklists stored as skills. A colleague types /complaint and gets walked through your standard process.
  3. Market monitoring. Once a week the agent browses competitor sites, flags changes in offers and pricing and logs them in a table.
  4. IT operations. SSH backend on a staging server, log analysis, a proposed fix — with every risky command waiting for approval.
  5. Content and search visibility. Article briefs, product description drafts, keyword summaries pulled from Search Console over MCP, with a human editing and signing off. See how this connects to our SEO and GEO work.

Security: what to configure before the agent touches company systems

An agent with a terminal and a messaging account holds real privileges. The Hermes security model has eight layers: user authorisation, dangerous-command approval, file-write safety, container isolation, MCP credential filtering, context-file scanning, cross-session isolation and input sanitisation. Configuration is what actually protects you.

Setting Recommendation Why
Gateway access Per-platform allowlists (e.g. TELEGRAM_ALLOWED_USERS) or DM pairing; never GATEWAY_ALLOW_ALL_USERS=true With no allowlist, everyone is denied by default — keep it that way
Terminal backend docker or ssh rather than local The local backend has no isolation; the Docker backend drops all but a few Linux capabilities and caps the process count
Command approval approvals.mode: smart or manual; never off or YOLO mode in production No answer within 300 seconds counts as a denial
Unattended jobs Leave cron_mode: deny Dangerous commands from scheduled jobs are refused by default
Skills skills.write_approval: true, lock critical skills Out of the box, the agent edits skills without review
Secrets ~/.hermes/.env at mode 600, outside version control Hermes strips KEY/TOKEN/SECRET variables from child processes, but that is no substitute for hygiene
Network Domain blocklist, SSRF protection left on The agent should not reach internal services or cloud metadata endpoints
Operations Run the gateway as non-root, watch ~/.hermes/logs/, run hermes update on a schedule From the production checklist in the docs

The docs are candid about two limits. Protected-path blocks (~/.ssh/, .env files) are defence in depth: the terminal tool runs as the same OS user and can still reach those files through shell commands. And prompt-injection detection in context files is heuristic, not semantic. Your real boundary is a container or a separate machine, plus least-privilege accounts for everything the agent can reach.

Data protection and GDPR

The project FAQ states that Hermes collects no telemetry and stores conversations, memory and skills locally in ~/.hermes/. API calls go only to the model provider you configure. If the agent processes personal data, that provider is your processor, so you need a data processing agreement and clarity on where data is handled. The alternative is a local model, where nothing leaves your infrastructure.

How to roll out Hermes Agent: step by step

  1. Pick one process with low risk and a measurable outcome, such as a weekly report or ticket triage.
  2. Provision a dedicated machine — a VPS or on-prem Linux server with a non-root user. Installation is one command: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash (Linux, macOS, WSL2), or a PowerShell script on native Windows. For staff laptops there is the Hermes Desktop installer.
  3. Connect a model. The fastest route is hermes setup --portal: one OAuth login covers a model plus four tools (web search, image generation, text-to-speech and a browser). You can also bring your own OpenRouter or OpenAI key, or a local endpoint; the docs use 64,000 tokens as the example minimum context window.
  4. Define identity and rules in SOUL.md (role, tone, hard limits) and in project context files such as AGENTS.md.
  5. Set up the gateway: hermes gateway setup, allowlists, then hermes gateway install, which registers a systemd service on Linux, a launchd agent on macOS or a Task Scheduler task on Windows (gateway docs).
  6. Load your procedures as skills with /learn, review them and lock the ones that must not drift.
  7. Add integrations over MCP with the narrowest permissions possible — read-only accounts wherever you can.
  8. Switch on cron jobs and watch for 2–4 weeks: logs, token spend, output quality, how often approval is requested.
  9. Pin the version. 0.x releases land often, so upgrade only after testing on a copy of the environment.

Already running OpenClaw? Hermes ships a migration path: hermes import-agent openclaw --dry-run previews what will move across (SOUL.md, memories, skills, command allowlist, messaging settings).

How much does Hermes Agent cost?

The licence is free; you pay for infrastructure, model usage and people’s time. Nous Research’s comparison page also mentions a hosted Hermes Cloud option, though company deployments usually run on their own server.

Cost item What drives it How to keep it down
Server Terminal backend, number of users, API vs local model (local needs a GPU with enough VRAM) A small VPS plus an API model; Daytona and Modal backends hibernate when idle
Model tokens Model choice, context length, cron frequency, number of subagents A cheaper model for auxiliary tasks, the built-in one-hour prompt cache for Claude models, execute_code instead of many turns
Implementation Number of integrations and skills, security requirements Start with a single-process pilot
Maintenance Release cadence, skill reviews, monitoring A pinned version and a monthly configuration review

Limitations people rarely mention

  • It is 0.x software. Configuration and behaviour can shift between releases.
  • Quality tracks the model. A weak local model plans worse and loses the thread on long tasks.
  • Memory is deliberately small. 2,200 characters is not a knowledge base — keep documents in skills or connect them over MCP.
  • Autonomy cuts both ways. The agent can be confidently wrong, so anything financial, legal or customer-facing needs a human sign-off.
  • External content is untrusted. Web pages and emails can carry hidden instructions, and heuristics will not catch all of them.

Wondering whether Hermes Agent fits your processes? At OrbitBuilder we build AI agents on Hermes and connect them to n8n automations. Get in touch and we will propose a single-process pilot with isolation and access controls in place from day one.

Frequently asked questions

Is Hermes Agent free?

Yes. Hermes Agent is released under the MIT licence, so you can use it commercially without licence fees. You pay for the server it runs on and for model usage with whichever provider you choose. With a local model there is no API bill, but you need hardware with enough GPU memory.

Can Hermes Agent run on a local model without sending data to the cloud?

Yes. Hermes supports custom endpoints including Ollama, vLLM, llama.cpp and SGLang. According to the project FAQ it collects no telemetry and sends requests only to the provider you configure, so with a local model your data stays on your own infrastructure. Make sure the model has a large enough context window.

How is Hermes Agent different from ChatGPT?

ChatGPT is a cloud service run by OpenAI that you use through its app. Hermes Agent is software you install on your own machine, connect to any model, and give access to a terminal, files and tools, then talk to through messaging apps. It also keeps its own memory and skills as files you control.

Does Hermes Agent run on Windows?

Yes. The docs cover a native Windows install via a PowerShell script as well as WSL2. Less technical users can use the Hermes Desktop installer for Windows and macOS. The gateway can be registered as a Task Scheduler task that starts at logon.

Is Hermes Agent safe for business use?

It can be, if configured properly. The essentials are user allowlists on the gateway, a Docker or SSH backend instead of local execution, smart or manual command approval, and reviewing skill changes before they go live. Give it access to company systems only through least-privilege accounts.

How does Hermes Agent learn?

It does not retrain the model. It writes notes to small memory files, searches past conversations stored in a local database, and creates or patches skills, which are procedure files. All of this lives in ~/.hermes/ as files you can read, edit and version.

Sources

  1. Hermes Agent — dokumentacja (Nous Research)
  2. Hermes Agent docs — Persistent Memory
  3. Hermes Agent docs — Skills System
  4. Hermes Agent docs — Security
  5. Hermes Agent docs — Messaging Gateway
  6. Hermes Agent docs — Features Overview
  7. Hermes Agent docs — FAQ
  8. Hermes Agent — Compare
  9. GitHub — NousResearch/hermes-agent
  10. GitHub — NousResearch/hermes-agent releases
  11. The Decoder — Nous Research releases Hermes Desktop (3 June 2026)

Need help with this?

An AI agent is more than a chatbot: it plans steps, uses tools, remembers context and works in your channels. We deploy agents built on Nous Research’s Hermes Agent and n8n.