Gateway 01 · Shared Console
BeamMeUp opens a real terminal, on Windows as on Linux, where you and an AI agent type into the same session, live. What the agent does, you see. What you type, it sees. No hidden window, no ghost process.
Code agents run inside a sandboxed shell whose standard input
is not a real terminal. The result: the slightest arrow-key menu, the
slightest y/n confirmation, the slightest interactive wizard
(eas credentials, npm init, an installer...) fails outright.
> npx eas credentials -p android
Input is required, but stdin is not readable.
PROCESS INTERRUPTED · NO INPUT POSSIBLE
BeamMeUp is not a server you start ahead of time, nor an MCP extension to
configure. The agent simply calls the program, the way it would call
git or npm. If the console isn't lit up, it
lights up, visible, right away.
Types on the keyboard, watches live, takes back control at any moment on the same session.
Opens sessions, sends commands, reads results, never without the console being on screen.
DIRECTIVE 01 No stealth mode. No execution without a visible window. Ever.
On Windows: PowerShell 5.1 and 7, cmd.exe, Git Bash,
every WSL distribution. On Linux: your login shell, then everything
declared in /etc/shells. Nothing is assumed, everything is scanned on the
real machine at every launch.
The system's real ssh/scp client is used as
is: your known_hosts, your key agent, no
homemade library to trust.
Name a session to find it again without memorizing a UUID, built for several agents running in parallel on the same machine.
Save a frequent command, replay it with one keyword, without ever storing a secret in it.
Grab a session's scrollback into a file, with or without ANSI codes, in full or just since the last command sent.
Remotely driven fullscreen, and on Windows a real screenshot of the console: the agent can literally look at what's on screen.
The agent sends a command and waits for it to actually finish: output and return code sent back directly, no more guessing a delay then re-reading.
List, read, write, rename, or delete a file on an SSH server without opening a tab: a minimal SFTP equivalent, via the same system client.
Duplicate a session, reopen the last closed one, close everything without quitting the console, or bring a specific tab to the front among several running at once.
Closing the window kills no session: a discreet icon stays in the notification area, with a menu to reopen the window, clear the tabs, or quit for good.
Off by default. Opt in to view and control a session from your phone over the network (your Tailscale network, for instance): you choose the bind address and the authentication, nothing is forced on you.
Visible is not the same as in front. No command from an agent raises the window: you keep working on whatever you were doing, and come back when you choose, with the right tab already selected.
A sidebar to pick a shell, see every session and replay a snippet, all with a click. You never have to go through the command line to follow or steer what the agent is doing.
Four transmissions are enough to open a session, work in it, then read back the result.
Survey the systems available on board
beammeup shellsOpen a named workstation (the identifier comes from the
previous command: pwsh5 on Windows, bash on
Linux)
beammeup open --shell bash --label my-workSend an instruction
beammeup send my-work "npm run build" --enterRead back the log, with no visual noise
beammeup read my-work --plain --lastOr wait directly for the command to finish, with its result
beammeup exec my-work "npm run build"The full tutorial (every command, every case) lives in the repo README.
Everything goes through your already configured system tools (SSH agent,
known_hosts). The only secret BeamMeUp keeps is the remote-access
token, if you turn that feature on, in a local file that is yours to delete.
The console runs as administrator so local sessions inherit its rights, without repeated prompts: a deliberate choice for a personal machine, not a shared server.
The command channel only listens locally, restricted to the account that launched the app. The screenshot port is local-only too, but any process on the machine can reach it: right for a personal PC, not for a shared one.
The one deliberate exception to local only, and it stays opt-in: you pick the bind address and the token, the console never decides for you.
The token is compared in constant time, and a server started without a token refuses requests that could come from a web page you merely visit (DNS rebinding). Remote access is plain HTTP, and says so: use it over Tailscale or a VPN, not an open network.
Windows 10/11 and Linux. Installers built and signed on every release, and published on winget (Windows) and apt (Linux).
Install with winget, the Windows package manager. The signed MSI and NSIS installers (Azure Trusted Signing) are also on the releases page.
winget install Breizhzion.BeamMeUpDebian package via the in-house repository, or a universal AppImage.
sudo apt install beammeupPaste this into your agent's instructions (an AGENTS.md, a
CLAUDE.md, a system prompt). A ready-made skill for Claude Code is
also in the repository, under .claude/skills/beammeup.
BeamMeUp is installed (command: beammeup). When a command needs a real
interactive terminal (a menu, a wizard, a y/n prompt, a password), do not
fail: open a visible session and work in it.
beammeup shells list the available shells
beammeup open --shell <id> --label my-work open a named session
beammeup exec my-work "<command>" run it, wait for output and exit code
beammeup send my-work "<text>" --enter type into the session
beammeup key my-work ctrl-c send a special key
beammeup read my-work --plain --last read what happened
The window is always visible to the human, who can step in at any time.
Full guide: https://github.com/bzhzion/beammeup#for-ai-agentsFull instructions (winget, apt repository key, AppImage, building from source) are in the repo README.
A program for Windows and Linux that opens real terminals (PowerShell, cmd, Git Bash, WSL, bash, zsh, SSH) in one window shared between you and an AI agent. The agent drives them through the beammeup command. You see everything live and can type in the same session.
No. BeamMeUp is not a server you start ahead of time. The agent calls the beammeup command the way it would call git, and the window opens, visible, if it is not already.
Any agent that can run a shell command, such as Claude Code or Codex. The README has a section written for agents, and the repository includes a ready-made skill for Claude Code.
An agent's sandboxed shell has no real terminal, so arrow-key menus, y/n confirmations and wizards such as eas credentials or npm init fail. In a BeamMeUp session they work, and a human can answer a password prompt.
It stores no SSH password or key, and the window is always visible. It runs as administrator by design, which suits a personal machine and not a shared server. Remote access from a phone is off by default and uses plain HTTP, so use it over Tailscale or a VPN. The security section and the privacy policy give the details.
No telemetry, no usage statistics and no outbound connection of its own. The privacy policy lists exactly what BeamMeUp writes to disk.
It is free to download and use personally, under the BZ-1.1 license. Manufacturing or commercial use for a third party requires a written commercial license.
On Windows, run winget install Breizhzion.BeamMeUp. On Linux, run sudo apt install beammeup after adding the repository, or use the AppImage. The installers are also on the GitHub releases page.
No. BeamMeUp is built for Windows and Linux, and has never been run on macOS.