Hermes Agent
The agent that grows with you
About Hermes Agent
Hermes Agent is an open source personal agent from Nous Research that runs as a terminal application or a native desktop app on macOS, Windows and Linux, and reaches you through Telegram, Discord, Slack, WhatsApp, Signal, email and Home Assistant from a single gateway. The problem it goes after is that most agents are tied to one surface. An editor plugin knows your codebase but not your calendar, and a chat assistant forgets everything between sessions. Hermes keeps one memory behind every surface it runs on, so context built up in the terminal is still there when you message it from your phone, and a skill it learned last week is available on all of them.
The project describes its core as a closed learning loop: it curates its own memory, creates skills out of work it has already done, and builds a persistent profile of the person it works for, with full-text session search and model-written summaries over past conversations. Skills follow the agentskills.io open standard rather than a private format. There is no model lock-in - it runs against Nous Portal, OpenRouter, OpenAI or your own endpoint, and you switch with a single command. Nous sells an optional Portal subscription that bundles monthly credits and catalogue access, and the honest caveat is that the tier prices are behind sign-in even though per-model token rates are published openly. Development moves quickly and the release is still pre-1.0, with a very large open issue count, so treat it as a project you follow rather than one you set and forget.
Installation is a one-line bash script on Linux, macOS and WSL2, a PowerShell one-liner on native Windows that bundles Git Bash, Python and Node.js without needing admin rights, or a documented manual path on Android through Termux. You then run it in a terminal that supports multiline editing, slash-command autocomplete and streaming tool output, or install the desktop app instead. Connecting a chat platform routes that surface through the same gateway, so Telegram, Discord, Slack, WhatsApp, Signal and email all reach the same agent with the same memory. Recurring work goes through a built-in cron scheduler described in natural language, and results are delivered to whichever platform you name. For anything long or parallel, the agent spawns isolated subagents with their own conversations and terminals so a large job does not consume the main context. Execution can be confined to one of seven sandbox backends depending on how much isolation the task deserves.
- •One Memory Across Every Surface - Telegram, Discord, Slack, WhatsApp, Signal, email, Home Assistant and the CLI all connect through one gateway to one agent and one memory
- •A Closed Learning Loop - Agent-curated memory, autonomous skill creation from completed work, persistent user profiles, and full-text session search with model-written summaries
- •Scheduled Automations - A built-in cron scheduler driven in natural language that can deliver its output to any connected platform
- •Delegation and Parallelism - Isolated subagents with their own conversations and terminals for parallel workstreams that cost the main session nothing
- •Seven Sandbox Backends - Local, Docker, SSH, Singularity, Modal, Daytona and Vercel Sandbox, with Modal and Daytona offering serverless persistence
- •Model Neutral by Design - Nous Portal, OpenRouter, OpenAI or your own endpoint, switched with one command, with skills following the agentskills.io open standard
Developers who want one agent reachable from wherever they already are rather than a separate window for each context, and who are comfortable running a fast-moving pre-1.0 project. The messaging fan-out is the differentiator: if you regularly want to hand work to an agent from your phone and pick it up at a terminal later, very little else is built that way. The sandbox choices matter to anyone letting an agent run commands against real systems, since local, Docker, SSH and remote serverless are meaningfully different risk levels. It is a poor fit if you want a coding agent bound to your editor, since that is not what this is, and a poor fit if you want something stable and pre-packaged - the release is pre-1.0, the surface moves, and the open issue count is enormous.
Pricing
$0 - $200/mo
- Free$0/mo
- Plus$20/mo
- Super$100/mo
- Ultra$200/mo
From the vendor pricing page, 2026-09-15
