ShellTeam — a cockpit for the coding agents on your VPS

the honest part

SSH + tmux works.
Until you stand up.

A VPS with coding agents on it is the best dev machine ever made: always on, real Linux, your keys, your tools. We're not here to replace that — ShellTeam never touches your shell, your dotfiles, or your agent configs.

It adds the missing layer on top: a cockpit you can open from a laptop, a tablet, or a phone on the train — where agents keep working after you close the lid, where their output is a page you scroll instead of a scrollback you fight, and where everything they make is one tap away from your eyes.

what changes when the terminal gets a cockpit

watch three agents at oncetmux panes → parallel slots

check on a build from bedimpossible → your phone

see the app an agent builtport juggling → live URL

agent stuck mid-task?kill & retype → steer, rewind, fork

wrong agent for the jobstart over → switch mid-conversation

describe a UI bugprose → paste a screenshot, or say it

what you get

One box. Four agents.
Zero footprint.

▸Agent-agnostic cockpit

Claude Code, Codex, Antigravity and OpenCode in one UI — stream output, interrupt, compact, switch models, run parallel tabs. Your subscriptions, not our markup.

▸Portable sessions

Hand a running conversation from Claude Code to Codex from a dropdown — the new agent continues with full context, in its own native session format. No other tool does this.

▸A URL for everything

Every file on the box is a link; every port an agent serves becomes you-3000.yourbox. An agent says "done" and you're looking at the result, from any device.

▸Phone-grade, for real

Voice input, paste-a-screenshot, readable diffs and tables on a 6-inch screen. Built by someone who actually drives production work from a phone.

▸Zero footprint

ShellTeam never writes to ~/.claude, ~/.codex or your gitconfig. Its layer is composed at launch-time and pinned by contract tests. uninstall.sh removes everything.

▸Opt-in superpowers

Modules, off by default: a shared browser your agents can drive, nightly "dreaming" that distils each day's sessions into per-project knowledge, 500+ app connections over MCP.

whose computer is it, anyway

Your box. Your keys.
Your agents.

01No vendor backend. Nothing leaves your VPS. Your LLM keys are proxied on-box so browser and tools never see them.
02Subscription-first. Log in with the Claude or ChatGPT plan you already pay for — API keys are the fallback, not the default.
03Native, not containerized. Three systemd --user services. No Docker, no Kubernetes, no 6 GB of images. Installs in about a minute.
04AGPL-3.0 and inspectable. Single-user by design — the code is small enough to actually read before you run it.

⚿ security, stated plainly

ShellTeam adds zero new agent privilege — the agents already had your box; we add remote access for you, behind one strong token over HTTPS, HttpOnly cookies, and derived read-only credentials for content. There's no container theater and we say so: read SECURITY.md before you expose anything. Use a dedicated, disposable box.

install

Two commands.
About a minute.

Fresh Ubuntu or Debian VPS, as your normal user. The installer is idempotent — re-run it any time.

bash

$ git clone https://github.com/sebderhy/shellteam && cd shellteam $ ./install.sh # pure core — zero agent injection, zero Docker

--minimal · the defaultCockpit, terminal, file server. Nothing injected into your agents.

--full · the experienceAdds the assistant persona, the shared browser, and nightly dreaming.

--public / --domainA real HTTPS URL via sslip.io or your own domain — or keep it on Tailscale.

SSH + tmux works.Until you stand up.

One box. Four agents.Zero footprint.