Hermes Agent vs Odysseus: two open-source AI tools, two different jobs

What each one is (and which Odysseus we mean)

The name is shared with the Greek hero and with a few unrelated repositories and forks. The project this page covers is the open-source AI workspace whose GitHub repository is odysseus-dev/odysseus. Its README describes it as “a self-hosted AI workspace for chat, agents, research, documents, email, notes, calendar, and local model workflows”. The repository was created on 31 May 2026, so it is a young project and it moves fast: its default branch is dev, and the setup guide warns that this branch may be unstable and points to main as the more curated one. The GitHub repository has no tagged releases; versions show up as Docker image tags, and the README’s own example pins 1.0.2.

Hermes Agent is the open-source agent from Nous Research. It is one agent with a terminal interface, a desktop app, a web dashboard and a messaging gateway. It keeps notes about you, writes reusable skills, runs scheduled jobs, and can use tools such as a browser, files and MCP servers. Its centre of gravity is that you talk to it from wherever you already chat.

Put simply: Odysseus is a place you go to, and Hermes is an assistant that comes to you.

Where each one runs, and what stays on

Odysseus is a server application. The recommended install is Docker, and the README’s quick start is a clone, a copied .env file and docker compose up -d --build. A native install needs Python 3.11 or newer, and its model-serving tool also wants tmux. You use it through a browser. By default the web interface binds to the local machine only; you change a setting to open it to your network, and the project’s security notes say to keep authentication on and treat it like an admin console, because it has shell access, file uploads and email integrations.

On Apple Silicon Macs, the setup guide says Docker cannot reach the GPU, so local models served from Docker run on the CPU; for GPU use there it recommends a native install. Odysseus has task scheduling (its README lists scheduled agent tasks), so it can do work on a timer as long as the server is running.

Hermes requires Python 3.11 to 3.14, or you can use the desktop packages for macOS and Windows. For always-on behaviour it uses the messaging gateway: one background process that connects your chat apps, keeps sessions, runs scheduled jobs and delivers voice messages. That process only answers while the machine it runs on is on and online. See running Hermes in Docker and the Hermes system requirements for the details.

Chat apps: the biggest difference

Hermes documents more than twenty places to talk to the same agent, including Telegram, Discord, Slack, WhatsApp, Signal, SMS, email, Matrix, Mattermost, Microsoft Teams, LINE, Feishu/Lark, WeCom and Weixin. It can also act as an OpenAI-compatible server for other front ends. Setup guides: Telegram, WhatsApp, Discord and WeChat.

In the Odysseus repository we did not find adapters that let you message the agent from Telegram, WhatsApp or similar apps. What it has instead is a phone-pairing bridge for its own companion clients on your local network, a Discord webhook integration (for posting out, not for chatting in), and ntfy push reminders. If your goal is “message my assistant from my phone like a contact”, Hermes is built for that and Odysseus is not.

Memory, skills and tools

Both projects give the agent memory, skills and MCP support, so on paper they overlap a lot. The shape differs.

• Hermes memory is deliberately small: two files, MEMORY.md (about 2,200 characters) and USER.md (about 1,375), loaded into the prompt at the start of a session and edited by the agent itself, plus search over past sessions. Heavier memory is available through optional memory providers. More in Hermes Agent memory.
• Hermes skills are on-demand documents in the open agentskills.io format. The agent can write new ones from what it learns, and you can install more from the Skills Hub. See Hermes Agent skills.
• Odysseus lists memory, skills, MCP, files and shell among its agent features. Its skills have a lifecycle in the interface (built-in, approved and draft) with automatic audits, and its documentation says this approach was adapted from Hermes’ progressive-disclosure design, without copying code.

Where Odysseus is clearly broader is the surrounding workspace. It ships a document editor with AI edits, an IMAP/SMTP email client with triage and reply drafts, notes, todos and a calendar with CalDAV sync, multi-step Deep Research that writes a report, a blind side-by-side model Compare mode, an image editor, and two-factor login. Hermes has a web dashboard and a desktop app, but not a built-in editor, mail client or calendar of this kind; it expects you to get such things through chat, tools and skills.

Models: hosted providers and local tooling

Hermes works with a long list of hosted providers, and you can set fallbacks and routing. See Hermes with Ollama or with Claude.

Odysseus also connects to hosted providers (its specs cover Anthropic, Google, Groq, DeepSeek, Azure OpenAI, Bedrock and many more) and OpenAI-compatible servers. Its Cookbook scans your hardware, recommends models that fit, downloads them and serves them through llama.cpp, vLLM or SGLang from inside the app, with GPU setup helpers for NVIDIA and AMD. Hermes also has built-in local-model tooling: it recommends models for your hardware, downloads models and a managed llama.cpp engine, and starts the local server. It can also connect to separately hosted endpoints such as Ollama and vLLM.

One caution from the Odysseus roadmap itself: it says agent mode is heavy for small local models because tool schemas, skills and memory can use up the context window, and it lists Cookbook reliability across different machines as high-priority work.

Licence and what you maintain

Hermes Agent is released under the MIT licence. Odysseus is AGPL-3.0-or-later, as its README and licence file state. For personal use the difference rarely matters. If you plan to modify the software and offer it to other people over a network, the AGPL requires you to make your modified source available to those users; MIT does not. Check with a lawyer if that is your plan.

Either way you maintain it yourself when you self-host: updates, backups, the machine, the network exposure and the provider keys. Odysseus adds some upkeep because it bundles several services (its guide mentions ChromaDB, SearXNG and ntfy alongside the app) and its roadmap lists fresh-install testing and an integration audit as open work. Hermes has its own moving parts: the gateway, per-platform credentials and updates. See the comparison of self-hosting and managed hosting.

Pick this if

Pick Hermes Agent if you want an assistant you message in Telegram, WhatsApp, Discord or a similar app, that remembers you, learns skills, and runs scheduled jobs in the background. It is also the one to choose if you want a terminal or desktop agent that works on code and files.

Pick Odysseus if you want one private browser workspace for writing, email, notes, calendar and research, you like comparing models side by side, or you want help picking and running local models on your own GPU. You are comfortable running a Docker stack and treating it as an admin console.

Can you use both? They are separate programs with separate data. We have not tested connecting one to the other, so this page does not claim that it works. Nothing stops you running them side by side for different jobs.

Hermes Agent on ReadyHermes

ReadyHermes runs your own Hermes Agent on our servers and keeps it on, so it keeps answering and running scheduled jobs while your computer is off. A new assistant is ready in about 10 seconds, with nothing to install. The storefront connects Telegram, WhatsApp, Discord, WeChat and Lark. It does not offer Odysseus.

You can use the AI included with the service or add your own key from a provider such as OpenRouter, OpenAI, Anthropic, Google Gemini or DeepSeek, which works during the free trial too. Changes that restart the assistant take about a minute. The free trial lasts one hour; see managed Hermes Agent hosting for what is included.