Argus

Watch your AI agents work. From anywhere. They already run in tmux — Claude Code, Codex, Gemini, a script that takes two hours. Argus puts those sessions, the files they are writing and the report they just produced in one browser tab, and the same thing on your phone. Nothing to install inside the agent: if it runs in a terminal, it works.

• Getting started — install it, run it, configure it
• Reaching it from outside — Tailscale, SSH tunnels, a reverse proxy, and the case where the network only goes one way
• Starting an agent — a shell or an agent, in a folder or a fresh git worktree, with its first instruction already typed
• Sessions and the terminal — attaching, typing, copying, sizing
• Files — browsing, editing, uploading, pasting screenshots
• Documents — markdown, PDF, Word, logs
• Desks and windows — workspaces, layout, the link tray
• Teams — agents taking turns on one goal, with a check between turns, directed by the server so it goes on with the browser shut
• Two agents on one job — the older pair, now replaced by Teams
• Notifications — being told when it has finished, or when it wants you
• Claude, Codex, Gemini compared — which agent rings, which can say where it is (what an agent may do in Argus is What an agent can do)
• Keyboard shortcuts — the keys, and how to change them
• On a phone — the drawer, the window sheet, and what a thumb can reach
• The machine — health, disks, listening ports
• tmux configuration — editing it from the app, safely
• Settings and languages
• Security — what protects it, and what does not
• Troubleshooting — the things that surprise people
• FAQ — the questions people actually ask
• Everything it does — the whole catalogue, one page, searchable
• What an agent can do — the table: by itself, with your tap (Do it / No), and not at all
• An agent driving Argus — a key of its own for an agent, and the calls that let your agents ask you, ring, and hand work to each other
• The Argus plugin — one install for Claude Code or Codex: hooks that say what the agent is doing, and MCP tools to use Argus
• Orchestrating several agents — three worked examples: N attempts and a judge, N referees and an editor, three roles on one repository
• Writing an orchestrator — the reference for argus_orchestra.py, when a team is not enough and you want a script of your own
• The API — the browser is one client of it, not a privileged one, and Swagger on your own machine
• API reference — every route, generated from the code
• Development — the tests, the three scripts, the vendored code

What it is

A small web app served by the machine your work runs on. It attaches to the tmux sessions you already have, the way another terminal window would, and puts a filesystem, a document viewer and the machine's own health next to them.

It is not a tmux replacement: kill Argus and every session carries on. It is not multi-user: one person's work on one machine (with a token per device, and narrower keys for agents and boards). And it is remote shell access, so read Security before putting it anywhere but a trusted network.

What it is for

An editor is built for the hours you spend writing code. More of that work is now done by an agent running for two hours in a session, and what is left for you is watching it, unblocking it when it asks, and judging what came out — a job an editor is not built for, and above all a job you often want to do from somewhere else.

In a hurry, or stuck: the FAQ answers the questions people actually ask — whether it can go on the internet (no), what happens when you log out, what the container is for, and what it sends anywhere (almost nothing).

---

Several machines? Panoptes is the board: every machine on one page, which sessions are on each, and which one is waiting for you. It holds one read-only key per machine and never gives a browser any of them. Argus does not depend on it — turn it off and every machine carries on alone.