Argus lets you leave AI agents running on a rented server, with spending limits

An agent you can leave alone

As github.com reported through a Show HN post (a Hacker News section where makers present their projects), Argus is a self-hosted system for agents that keep working for days or weeks. Self-hosted means you install it on a server you control, not on someone else's cloud service. The current release is 0.7.3.

The basic unit is a project: one long-running agent with its own folder, chosen AI model, instructions, memory and spending limit. Argus ships with starting templates for website work, research, reports and devops (the upkeep of servers and software).

You can give a project one-off tasks or put it on a schedule. The schedules are meant to fire exactly once even if the server restarts, and results arrive in your chat.

The guardrails

Most of what Argus offers is about keeping an unattended agent from doing damage. Here is how the project describes it:

• Spending gatekeeper. Every run goes through one component that checks limits before the run starts and again at each step. Model prices update automatically, and the /panic command stops everything in under five seconds.
• Approvals. Each project gets its own permissions for reading, writing, running shell commands and using the web. Risky actions pause for your answer, and if you never answer, the action does not happen.
• Sandbox. Using a Linux kernel feature called Landlock, the agent's commands are locked inside that project's folder. A command named /undo rolls the folder back to how it looked before a run.
• Runaway protection. A run is never paused so it can ask for more time. It ends when it hits its budget, its step limit, or a loop that ordinary code detects. You can also add a second model to watch for loops.

The chat commands such as /status and /stop are handled by regular program logic, not by the AI. That means they still work if the model is confused.

Git, files and add-ons

Projects can clone, pull and push code repositories through Argus. Access tokens are held per project and per remote, and the agent never sees them. Before a push, Argus shows the commits and asks first.

The web page accepts uploads up to 2 GB per file, lets you download any folder as a zip, and turns files too large for chat into one-time links.

Add-ons come from git and are pinned to a version, then switched on per project. The first-party set covers Word, Excel and PowerPoint files, a headless browser limited to the public internet, video and audio processing, diagrams, PDF handling and image editing.

Setting it up

You need a Linux server and an API key from an AI model provider. Telegram is optional. There are two install routes: Docker (recommended) or a native install managed by systemd, the standard Linux service manager. In both cases the installer asks for what it needs and gives you a setup link.

The web interface listens only on the server itself, on port 3091. To reach it from your laptop, use Tailscale (a private-network tool) or an SSH tunnel. That keeps the control panel off the open internet.

How much to trust it

The project is frank about its youth. It runs every day on its author's own server, and the repository says it has more than 2,000 automated tests, including end-to-end checks of installs, upgrades and the sandbox. But nobody else has used it in production yet.

It also sits on top of DeepSeek Harness, a framework that is still at release-candidate stage (0.2.0-rc.2). On Hacker News the post had 1 point and 1 comment when we looked, so the community has not vetted it. Our newsletter hosts have not run it either.

The license is the Sustainable Use License, the same one n8n uses. You can run and modify Argus for yourself or your company, including commercial work, but you cannot sell it or offer it as a hosted service.

What comes next

The roadmap lists release 0.7.4 with email (IMAP) as a way to talk to your agents. After that come Slack or Discord, team roles, semantic memory (searching by meaning, not exact words) and usage metrics. The author says an issue describing your own use case is the most useful feedback.

One caution on costs: the project's price data comes from public datasets, and it says your provider's invoice is the final word on what you owe.