Multi-step task execution
Give Hermes a goal, not a prompt. It breaks down the task, executes each step — code, web research, file operations — and adapts as it goes. Results are delivered, not just discussed.
Applications → Hermes Agent
An AI that acts, not just chats.
Hermes Agent is an autonomous AI workspace powered by NousResearch's Hermes. Unlike a chatbot, it can plan and execute multi-step tasks: write and run code, browse the web, manage files, execute shell commands, and interact with your ReefOffice apps — all through a secure web interface on your own server.
What Hermes Agent does
Hermes Agent is an autonomous AI agent designed for task execution rather than conversation. You give it a goal — "update this CRM record", "research competitors and write a summary" — and it plans, executes, and iterates until the task is done. It runs on your own ReefOffice server, with access to local models (AI plans) or your own AI provider key (Office plans). The ReefSkills pack teaches it your stack from day one.
From conversation to execution
Web browsing & file management
Hermes can browse the web for research, download and process files, and manage your server's file system. All actions are logged and reviewable in the audit trail.
ReefSkills pack
Every Hermes instance ships with a curated skill pack that teaches it the ReefOffice stack — how to query Grist, trigger automations, read Paperless documents, and interact with your apps. Knowledge is built in, not learned through trial and error.
Skill learning modes
Choose how Hermes evolves: in default mode, self-written skills reset on redeploy for predictable behaviour. In tracked mode, skills are version-controlled and auditable via your Forgejo instance.
Sovereign by default
Hermes uses your local Ollama models by default. Nothing leaves your server unless you explicitly configure an external AI provider key. Your data, your tasks, your infrastructure.
Skill marketplace
Hermes ships with a growing library of pre-built skills — document summariser, invoice data extractor, email drafter, calendar manager and more. Install skills from the dashboard or write your own in Markdown.
How it's deployed
Two containers, one secure agent workspace
Hermes runs as two Docker containers on your ReefOffice server: the agent backend (port 8642) and a web UI (port 8787). The web interface is protected by Authentik SSO. All agent actions are logged. The agent requires Ollama for local inference (AI plans) or a client-supplied AI provider key (Office plans). Persistent data lives at /opt/hermes/data, including skills and audit logs.
Connected to your workspace
Hermes integrates with the full ReefOffice stack. It can query Grist tables, read Paperless documents, trigger Activepieces automations, and manage files in Nextcloud. The optional Forgejo integration provides a version-controlled audit trail of the agent's self-learning. Everything is scoped by design — skills call on-box RAG and automation shims, never touching business logic directly.
Common questions
How is Hermes different from Open WebUI? +
Open WebUI is a ChatGPT-style chat interface for conversation and document Q&A. Hermes is an autonomous agent that executes multi-step tasks — writing code, browsing the web, managing files, running commands. Use Open WebUI to talk to your documents; use Hermes when you need something done.
Does Hermes need internet access? +
By default, Hermes runs fully offline using local Ollama models. Web browsing and external API access are available only when you provide an AI provider key or configure them explicitly. You control what leaves your server.
Can I review what Hermes did? +
Yes. Every action Hermes takes is logged in the audit trail, accessible through the web UI. In tracked skill learning mode, skills are additionally version-controlled in Forgejo, giving you a browsable, reviewable history of how the agent evolved.
Ready to hand tasks to your AI agent?
Book a demo and see how Hermes Agent on ReefOffice can execute multi-step tasks autonomously — on your own private server.