
Open-source autonomous agent from Nous...
A chat window that only answers questions is not the job this project describes. It stores procedures as reusable skills, compatible with agentskills.io, that you can carry, share, and revise on later tasks. Full-text search recalls across sessions, and Honcho builds a longer model of how you work. Messaging is not limited to a terminal on one desk: Telegram, Slack, Feishu, DingTalk, and WeCom can share one gateway. A scheduled job can deliver back to a connected platform. The runtime can be the local machine, Docker, SSH, Daytona, Singularity, or Modal. It fits developers who will install it themselves, bring their own model provider, and want the agent to keep their working habits.
Hermes Agent is an open-source autonomous agent from Nous Research under the MIT license, and installing and using the agent itself does not require a fee. It creates skills from work it has done and improves those skills later. Cross-session recall uses full-text search, and Honcho keeps a running model of the user. One gateway reaches more than a dozen messaging platforms, including Telegram, Discord, Slack, Feishu, DingTalk, and WeCom. It includes scheduled jobs and MCP connections. Runtimes include the local machine, containers, and remote hosts, and it can sleep when idle. The official install script supports Linux, Mac, and WSL2, and you configure your own model provider before the first run.
Difficulty: Advanced
Skills created and revised in use
After a task it stores the procedure as a skill, improves it in later use, and stays compatible with agentskills.io so it can be carried and shared.
Memory across sessions
Full-text search recalls across sessions, and Honcho keeps building a model of the user over conversations.
One gateway for many chats
Telegram, Discord, Slack, Feishu, DingTalk, WeCom, and other platforms connect through one gateway.
Schedules and several runtimes
Built-in schedules can deliver to a connected platform. Runtimes include local, Docker, SSH, Daytona, Singularity, and Modal.
Drive it from a chat app
A developer connects the self-hosted agent to Telegram or Feishu, assigns work in chat, and gets the result back.
Turn a check into a schedule
A repeating check becomes a scheduled job that runs and delivers the result to a platform already connected.
Keep a workflow as a skill
A procedure used once is stored as a skill, reused on the next similar task, and can be shared.
Yes. The documented steps are to run the install script, then a setup command that connects your own model provider, and only then start a conversation. The script supports Linux, Mac, and WSL2. It does not ship a model that works with no configuration. Without a model provider, the agent cannot run as the page describes.
The page describes skills as procedural memory that remains. The agent creates them during tasks, improves them in use, and can carry or share them through agentskills.io. Full-text search is a separate way to recall across sessions. Not every casual line becomes a skill; reusable procedures do. The files live in the local install.
No. The agent is described as open source, with no fee to install and use. The same page also promotes a separate image, video, and music API that bills per successful task and has its own key. Do not treat that API's billing as the price of the agent installer. API rates belong on the API page and are not listed here.
Real reviews and feedback from users