Self-Hosted AI Agent: What You Install, and What You Still Pay?
A self-hosted AI agent is the install you run yourself. Hermes Agent leaves the model bill separate. OpenMuse is a template, and Intelligence is a separate subscription.
A self-hosted AI agent is the program you run, with the model bill kept separate from the install. Hermes Agent is Nous Research's agent, released under the MIT license. You point it at a model provider you already pay, or at Nous Portal. OpenMuse is CopilotKit's personal-agent template, also MIT, and the page marks it alpha. The local sample runs without a model key. The connections that make it useful are added after.
OpenMuse is not Muse. Muse is Meta's hosted agent. This page is about the two you install.
Quick answer
| Hermes Agent | OpenMuse | |
|---|---|---|
| What you install | The program is free. Desktop builds are a beta preview for macOS 12 and newer, Windows 10 and 11, and Linux. You can also use the terminal, or Telegram, Discord, Slack, WhatsApp, Signal, or email. Cloud hosting is optional | Five commands: clone the repo, install with pnpm, copy the env file, then start the API on port 8787 and the web app on port 8081. That sample needs Node 24 and pnpm 11, and no model key, Google account, Docker, or Intelligence subscription |
| What you still pay | Portal plans, if you use them: Free is US$0, Plus is US$20, Super is US$100, and Ultra is US$200 a month. Those cards are model and tool credits. They do not print a price for the hosted agent itself. A provider you already use stays on that provider's bill | OpenMuse prints no price of its own. CopilotKit Intelligence is a separate subscription. Developer is free, Pro is US$39 a month, and Team is US$100 a seat a month |
| What remains when it stops | Stopping a local agent ends the process and does not delete saved conversations, memories, or skills. Tasks that need a running agent do not run while it is stopped | A memory can be edited or forgotten. If you connect Intelligence, threads are kept 3 days on Developer, 5 days on Pro, and 14 days on Team |
| Where the pages disagree | The paid cards say 200+ models. The Portal doc says 300+, and the last row of its table says 280+. Hermes 4, the 70B and 405B chat models, is discounted on the Portal and is not the model the agent loop should run | The product page says chat, the computer, Gmail, Calendar, and the other named surfaces run in the alpha today. The repository inventory still marks live Google acceptance, a networked computer, and live model testing as unfinished |
Hermes Agent: the install is free
The homepage FAQ says you can bring your own model provider, or connect Nous Portal for model and tool credits. Cloud hosting is optional. The four plan cards do not print a price for that hosted agent. The desktop preview on the same page is labeled a beta preview.
The cards are the Portal plans, and the credit figures are dollar amounts. Free lists free models only and US$0 in monthly credits. Plus is US$20 a month, with US$22.00 in credits, a 10% bonus, and a US$10.00 rollover cap. Super is US$100, with US$110.00 in credits and a US$50.00 rollover cap. Ultra is US$200, with US$220.00 in credits and a US$100.00 rollover cap. Plus is the column marked most popular.
The Portal is optional. An install that already uses OpenRouter, Anthropic, or another provider keeps that setup. The Portal is then one more provider, switched with /model during a session or with hermes model between sessions. If hermes portal info shows the model is currently another provider, that usage is not on the Nous bill.
The paid cards say 200+ models. The Portal doc titles the catalog 300+ frontier models, and the last row of its table says 280+ additional models. The same page says Hermes 4, the 70B and 405B chat models, is on the Portal at a heavy discount, and that those models are not the ones to run inside Hermes Agent. It calls them chat and reasoning models, not the tool-calling loop the agent uses. For the agent it points at other names, including Claude Sonnet 4.6, GPT-5.5 Pro, Gemini 3 Pro, and DeepSeek V4 Pro.
Paid Portal plans include the Tool Gateway. It routes web search, image generation, text-to-speech, and a cloud browser through Nous, and those calls are pay-as-you-use against the subscription. Some accounts also get a free tool pool. When it is there, it funds the FAL image models only. The Hermes Agent page keeps the plan lines and the model-count mismatch together.
The homepage says the agent learns your projects, writes skills, and never forgets how it solved a problem. The FAQ is narrower: stopping a local agent ends the running process and does not delete saved conversations, memories, or skills. Tasks that need an active agent do not run while it is stopped. Hosted agents are managed separately in Nous Portal.
OpenMuse: the sample is not the alpha list
The local sample is five commands, and the page says it needs no model key, no Google account, no Docker, and no Intelligence subscription. The next sentence says to add those when you want the real thing. apps/mobile is the shared iOS, Android, and web interface. apps/server is the API, with the task engine and the review step. apps/worker is a token-protected browser with a persistent Chromium profile. apps/computer is the optional Linux image.
The same page says these surfaces run in the alpha today: chat, the agent computer, activity, ideas, goals and tracking, documents, Gmail and Calendar, personal context, and AG-UI streams. Gmail and Calendar are described as full threads, drafts, and calendar changes. Every send, and every change to an event, waits for your review.
The repository's feature inventory says planned extensions are not claims of current support. The Linux computer is disabled until you configure it, and after that it has no network. It is not a graphical desktop and not a full virtual machine. Live Google acceptance, editing a repeating event, and any connector besides Google are in the unfinished column. Live testing of chat against a real model account is unfinished too. Ideas can be accepted, edited, or dismissed, and a memory can be edited or forgotten.
OpenMuse does not print its own prices. CopilotKit Intelligence is what the page says keeps the agent improving: it learns from the agent and proposes skills, and a skill is written in only after you approve it. The limits, if you connect it, are on CopilotKit's pricing page. Developer is free, for one developer, and keeps threads for 3 days, capped at 200, with 500 channel credits in total rather than a monthly refill. Pro is US$39 a month, keeps threads for 5 days, and gives 2,000 channel credits a month. Team is US$100 a seat a month, up to 5 seats, keeps threads for 14 days, and says self-hosting includes the database. Developer and Pro say a private-network deployment is the runtime only.
On that pricing page, Self-Learning is marked early access, on the Enterprise column. The OpenMuse page describes Intelligence proposing skills now. The OpenMuse page keeps both.
Which one to open
- You already pay a model provider, or you want the agent in a terminal or a chat app, and you can leave the Portal off. Start with Hermes Agent. The program is free. Read the 200+ line and the 300+ line before you trust a model count, and do not pick Hermes 4 for the agent loop. Their own Portal doc says so.
- You want a personal-agent repo you clone, and you can treat the alpha list as a target. Start with OpenMuse. The five-command sample is the part that runs without keys. Gmail, the computer, and Intelligence are the part you add, and the inventory still marks pieces of those unfinished.
- You want an agent that stays on someone else's computers after you close the laptop. That is a different job. Cue, Muse, Dots, and Grok Bot are that comparison. Muse on that page is Meta's product, not OpenMuse.
Hermes Agent and OpenMuse are on Seailife if you want the plan cards beside this split.