ZeroG Terminal is an Electron workspace manager for persistent terminal and SSH sessions on Linux and Windows, with smart features to help keep you productive. It is designed with AI tooling in mind: leave Codex, Claude Code, or other terminal-based agents running in persistent sessions, then return to them without repeating setup and resume commands.
ZeroG Terminal is a multi-session terminal and SSH workspace. It uses screen where available for resumable sessions and falls back to a direct shell PTY when screen is not installed.
- Multi-pane workspaces with stack, vertical split, horizontal split, and four-pane grid layouts.
- Draggable dividers between panes and beside the sidebar, so a split does not have to be an even one. Sizes are remembered between launches.
- A compact navigation rail for sessions, overview, settings, SSH connections, and opening a new local terminal, while keeping the sessions sidebar collapsible.
- Layout controls that restore a split, even up its pane sizes, and maximize or restore a pane in a predictable sequence.
- Maximize a focused pane and cycle between sessions without losing the other panes.
- Local sessions powered by Bash, PowerShell, WSL, and other supported shell backends; persistent sessions use
screenwhere available, with a process-only fallback when it is unavailable. A WSL pane starts in the distribution's own home directory rather than the Windows one seen through/mnt/c. - SSH sessions for hosts,
user@host, anduser@host:porttargets. - SSH configuration discovery from known connections, including remote
screensession discovery. In the sidebar's Connections tab, clicking a saved connection opens the connect dialog with the host filled in, and double-clicking it skips the dialog and opens the host in a new pane. - Reconnect to existing local or remote
screensessions from the Screens view. - Remote screen discovery for the selected SSH destination, including manually entered hosts, with refresh and visible failure states. Background discovery uses key/agent authentication; password-only hosts cannot be queried through the terminal connection.
- Remote screen attachment launched directly through SSH, preserving the account, custom port and SSH config alias, with authentication and host-key prompts left to the user.
- Shared ports over SSH, in a Ports view opened from the rail: forward a port on a remote host so it answers on this machine, or a port here so it answers on the remote. Each tunnel is its own SSH connection, so a host needs no terminal open first. Ports bind to loopback unless you widen them, and are remembered between launches.
- A directory browser inside a pane, opened from the folder button in its title bar: the pane splits, terminal on one side and the directories on the other, with a draggable divider. Double-click a folder to take the shell there. It works on SSH panes and on local ones, WSL included, follows a
cdyou type by hand, and is remembered per pane when you switch workspaces or relaunch. - An SFTP transfer panel, opened from the ⇅ button above the panes: local files on the left, the active SSH session's host on the right, with upload, download, new folder, rename, and delete. It connects to the host that session is already using and opens at the directory its shell is standing in, so a file can go straight to the project being worked on.
- Session history for reconnecting to sessions after a relaunch, with bounded structured history and no stored secrets.
- A ranked command palette on
Ctrl+Shift+R, in the spirit of McFly: commands you have run, ranked by directory, host, recency, frequency, whether they worked, and whether you picked them before. Off by default — it is the one feature that stores what you typed — and it refuses anything that looks like it carries a credential. - Workspaces for grouping sessions and quickly switching between projects or tasks. Drag the tabs beside the wordmark to put them in the order you want;
Ctrl+Shift+1…9follow that order, and it is remembered between launches. Right-click a tab to rename it, duplicate it, move it, or close it. Each workspace keeps its own panes, layout, focused terminal, and maximized pane, so switching to one restores the arrangement you left it in. Workspaces survive a relaunch: localscreenterminals reattach on their own, and SSH panes come back as ghost rows that reconnect when clicked, rather than dialling out to a host on startup. - Session overview, collapsible sidebar, keyboard shortcuts, and nine colour themes: Tokyo Night, Light, Dracula, Nord, Solarized Dark, Monokai, One Dark, Synthwave and High contrast. The palettes are matched by hand to those open-source themes.
- xterm.js terminal rendering with scrollback preservation while changing layouts.
- Voice input, either with Whisper ONNX inside the app through Transformers.js or through an OpenAI-compatible transcription server you point it at — on this machine, on the LAN, or hosted; transcribed text is typed into the selected terminal without automatic execution.
- A per-pane proceed button that sends a configurable phrase —
OK, proceedby default — for waving an agent on without typing the same reply again. - A settings panel for appearance, terminal behaviour, session defaults, and speech recognition, including a built-in recognition test.
- AI command suggestions from any OpenAI-compatible endpoint — Ollama, LM Studio, llama.cpp, vLLM, OpenRouter, or OpenAI itself — configured with a base URL, a model and an optional key in Settings. Ask what you want, get one command with an explanation, and approve it before it runs.
- Sandboxed Electron renderer, context isolation, disabled Node integration, and a narrow typed preload API.
- Safe argument-array handling and validation around SSH and
screensession operations. - Optional local MCP control for inspecting and restoring workspaces, and creating explicit project workspaces. It is loopback-only, bearer-authenticated, lease-controlled, and never executes terminal commands; see docs/mcp.md.
The project is particularly useful for terminal-based AI development workflows: start an agent in a persistent session, disconnect or suffer an interrupted connection, and reconnect later to see what it has done and continue working.
See the project walkthrough on YouTube.
ZeroG Terminal 1.0.0 was the first stable release. The current release is 1.1.0; the version history is tracked in versions.txt. The running version is shown beside the wordmark in the title bar, read from the app itself rather than written into the interface, so it is accurate in a packaged build too.
The GitHub Releases page provides a Windows x64 installer and portable executable plus a Linux x64 AppImage and Debian package for each desktop release. The npm package remains available for developers who prefer to launch ZeroG from Node.js.
Ctrl+Shift+C— copy selected terminal textCtrl+Shift+V— paste into the active terminalCtrl+Cremains the interrupt signal (not copy)Ctrl+Shift+N— new workspaceCtrl+Shift+1…Ctrl+Shift+9— switch to a workspace by positionCtrl+Shift+T— new local terminal in the current workspaceCtrl+Shift+R— ranked command history paletteCtrl+Shift+A— ask the AI endpoint for a commandCtrl+Shift+L— fold to a single pane, or back to the last splitCtrl+Shift+O— session overviewCtrl+Shift+B— toggle sessions sidebarCtrl+Shift+,— settingsCtrl+Shift+/— the help panel: what is here, and this listEsc— close overview / dialogs, cancel voice recording
Every one of these can be moved: Settings → Shortcuts lists them, and
pressing a new chord in a row rebinds it. A shortcut needs Ctrl and at least one
of Shift or Alt, so a shell keeps its own bindings — Ctrl+C interrupts,
Ctrl+R is still readline's reverse search, Ctrl+L still clears the screen,
and Ctrl+A still reaches screen or tmux. None of these shortcuts reaches
the shell in the pane that has focus, which was verified by asking bash to report
the control characters it receives.
Rebinding exists because nothing in the app can outrank the machine it runs on.
Ctrl+Shift+T was reported already taken on Windows 11; Windows itself uses Ctrl
and Shift together to switch keyboard layout when more than one is installed; and
some applications — Teams, for one — register chords globally while they are
running. Ctrl+Alt is usually clear, though on some layouts it behaves as AltGr,
so a chord you would type in a terminal is worth avoiding. Ctrl+Shift+1 … 9
and the clipboard pair cannot be moved.
The ? in the title bar opens the same help panel, which lists the shortcuts as
they are currently bound rather than as they shipped.
The Ports view, the SFTP transfer panel and a pane's directory browser have no
shortcut of their own; they open from the rail, the ⇅ button above the panes, and
the folder button in a pane's title bar. Esc deliberately does not close the
directory browser: it is part of the pane rather than an overlay, and Esc has
to keep reaching the shell for vi to be usable.
Clicking a link in a pane opens it in your own browser rather than in a window
of ZeroG. Hovering one first shows where it actually goes in the status bar,
which matters because a terminal hyperlink can be labelled with anything. Only
web links and mailto: addresses are opened; anything else says so in the
status bar instead, since a link in terminal output can name a scheme that would
start local software.
Selecting text with the mouse also copies it, and programs running inside a terminal can copy to the system clipboard themselves through the OSC 52 escape sequence — this is how TUI tools such as CLI coding agents, tmux and Neovim put text on the clipboard, including over SSH. Reading the clipboard through OSC 52 is refused, so a program on a remote host cannot see what you last copied.
Ctrl+Shift+R opens a search over the commands you have actually run, ranked so
the one you want is usually first: same directory beats same host, recent beats
frequent, something that worked beats something that failed, and something you
picked from this palette before beats something you merely ran. Typing matches a
subsequence, so gcm finds git commit -m, and the matched characters are
highlighted so it is clear why a row is there.
Enter puts the command on the prompt without running it. The command came out of a store rather than from your hands a moment ago, so pressing Enter is your decision — and it leaves room to change an argument, which is most of why anyone reaches for history.
Ctrl+R is untouched. Your shell keeps its own reverse search; McFly can rebind
that key because it is the shell, and ZeroG taking it would remove
reverse-search from every pane including remote ones this feature cannot see.
Recording is off until you switch it on in Settings under "AI & voice". It is the only part of ZeroG that stores what you typed, so it is opt-in rather than a default you have to find and disable.
Commands are read from the standard OSC 133 prompt marks a shell emits — the same marks VS Code, kitty, WezTerm and Windows Terminal use. If you already have shell integration from any of those, or from an oh-my-zsh plugin, ZeroG reads what is already there and you need to add nothing. Otherwise Settings offers a snippet per shell to paste into your rc file, and the panel says how many of your open panes are actually reporting marks, so "did that work" has an answer.
A remote host needs the snippet installed on the remote host: the marks come from the shell, and over SSH that shell is on the far side. A pane whose shell reports no marks records nothing, rather than guessing from the screen and recording something wrong.
Each entry holds the command, the directory and host it ran in, its exit status,
when it last ran, how many times, and how often you picked it. It lives in
command-history.json in ZeroG's profile directory, written with owner-only
permissions. "Forget all remembered commands" empties it and deletes the file.
A command that looks like it carries a credential is refused outright rather than
stored with the value masked — a masked entry still records that you set a
particular secret in a particular directory at a particular time. Refused shapes
include assignments to anything named like a token, key, secret or password;
--password, --token and --api-key flags; credentials inside a URL;
Authorization headers; curl -u user:pass; vendor-shaped tokens; and long
high-entropy words.
This is not a complete defence and is not claimed to be. A secret typed as a bare
argument to an unusual program, or piped in from echo, looks like ordinary text.
See SECURITY.md.
Drag the line between two panes, or the sidebar's right edge, to change how the space is shared. Sizes are remembered between launches and clamped so that no pane can be dragged down to nothing.
The dividers take keyboard focus as well: the arrow keys nudge one two percent at a time, and Enter or a double-click puts it back in the middle. One divider position is shared by every layout, so a split you set up in the vertical split is the same split you get in the four-pane grid.
The Ports button in the rail opens a list of shared ports, grouped by the host each one runs through. "Share port" asks for a host and a port, and the port is then reachable as though the service were running here.
Two directions are available under "More options". The default sends a port on the remote to this machine, which is what a remote dev server, database, or debugger needs. The other sends a port here to the remote, for a webhook or an agent on that host calling back. A row always names the side that listens first, so which way a tunnel runs is never left to be inferred.
A shared port answers only on the machine that binds it, unless you tick "Share
on my network" — which re-exports the service to whatever network that machine is
attached to, and is marked LAN on the row so it cannot be forgotten about. A
port sent to the remote is bound on loopback there by default; widening it also
needs GatewayPorts enabled in that host's sshd_config, and the row says so
when the server refuses.
Each tunnel is a separate ssh process, which is what makes closing one exact:
the cross stops that tunnel and nothing else. It also means a host that
authenticates with a password asks once per tunnel, in the panel, and nothing
typed there is stored. Hosts using a key or an agent are not asked at all.
Shared ports are remembered between launches and come back listed but not connected. Clicking one reconnects it — the app never opens a connection to a host on its own at startup.
Set an endpoint in Settings under "AI & voice": a base URL ending in /v1, a
model, and a key if the endpoint wants one. One field serves every provider,
because chat/completions is the request shape they all implement — a local
Ollama at http://127.0.0.1:11434/v1 needs no key at all. "Test connection"
asks the model for a token and reports what came back, and "Refresh" lists the
models the endpoint says it has. A key is stored encrypted by the operating
system, never in the settings file, and is only ever read in the main process:
it is not held in the window.
Suggest asks what you want, then returns a single command with an explanation. Only one command, ever — a reply naming several, or answering in prose rather than the structure asked for, produces no command to run and says so. The command is written to the pane you asked from, not to whichever pane is focused when the answer arrives.
By default the model is told only your shell, directory and host. Turning on "Send recent terminal output" also sends the tail of the focused pane, which is what lets a suggestion read the error you are actually looking at — and means terminal content goes to whatever endpoint you configured. The panel says which of those is happening.
While output is being sent, approval cannot be turned off. Terminal output can come from a remote host, and a host can print text shaped like an instruction; a command chosen downstream of that must be read before it runs. The output is sent as clearly delimited data with the delimiter stripped out of it, and the answer is only believed if it arrives in the exact structure requested — but neither of those is trusted to hold on its own, which is why the approval step is not optional in that configuration.
The ⇅ button above the panes opens a two-pane transfer panel: this computer on the left, the active SSH session's host on the right. It is only available while an SSH session is selected, and the button says why when it is not.
The connection is made with the system sftp client, so it uses the same
~/.ssh/config, agent, keys, and known_hosts as the terminal beside it. A
password, a key passphrase, or an unknown host key is asked for inside the
panel — ZeroG never answers a host-key question on your behalf, and the
fingerprint is shown with the question. Nothing typed there is stored.
The remote side opens at the directory the terminal's shell is currently in,
where that can be known without disturbing the session. ZeroG reads it from
OSC 7 — the sequence a shell emits to report its directory — and otherwise from
the path in the prompt; it never types pwd into your session to find out. When
neither is available the panel opens at the login directory.
Select files with a click, or several with Ctrl-click or Shift-click, then Upload or Download.
Double-click a folder to open it, or type a path into the folder box. New folder,
rename, and delete act on one selected item; deleting asks first, and a folder
must be empty, so a single click can never remove a tree. Remote folders can be
downloaded recursively with the system sftp client; uploads remain file-only,
so a local folder is never copied accidentally as a different job.
Filenames containing quotes, backslashes, or the wildcard characters * ? [ ]
are refused with a message rather than acted on. The sftp client re-reads its
own arguments through a glob pass, and there is no encoding of those characters
that is provably correct for every command — being approximately right about
which file to delete is not good enough.
Right-clicking a tab offers Duplicate, which copies the arrangement rather than the sessions: the layout comes across, and each SSH pane comes across as a pane waiting to be reconnected, so clicking it opens a new session to the same host. A local pane is not copied — its shell is a process on this machine, so a copy would be a different terminal rather than a duplicate of anything, and the status line says how many were left behind rather than letting you notice a pane missing.
The folder button in a pane's title bar splits that pane: the terminal on one side, the directories where its shell is standing on the other. Drag the divider to change the balance, double-click it to centre it, or use the arrow keys when it has focus. A pane remembers whether its browser was open and where the divider was, per workspace and between launches.
Single-click a folder to look inside it without moving the shell.
Double-click it to take the shell there — ZeroG types the cd into the pane,
which is deliberately visible: you asked for it, and seeing the command is how
you know what happened. The button in the bottom-right corner does the same for
the directory already on screen, which is how you retry after a cd that was
refused because the pane was busy, or catch the shell up after browsing around.
.. is always the first row, except at a root the shell cannot go above.
A single click waits a moment before it acts, because it has to find out whether it is the first half of a double-click: browsing immediately would replace the row under the pointer and the second press would land on a different one.
The browser opens on the directory the pane's shell reports, and on the login directory before it has reported anything — a pane that has just connected has run nothing, so there is nothing to read yet. That login directory comes from the connection itself, never from a command typed into your session.
Where it lists from depends on the pane. An SSH pane lists over an SFTP
connection to its host, opening one if there is not already one open — the same
client, ~/.ssh/config, agent and known_hosts the transfer panel uses, and the
same connection when both are looking at the same host. If the host asks for a
password, a key passphrase, or a decision about its host key, the browser says
so and the transfer panel (⇅) is where you answer: it shows the question with
the key fingerprint beside it, and ZeroG never answers one on your behalf. The
login directory that connection reports is also how a ~ in the pane's prompt
becomes a path. A native local pane lists the filesystem directly. A WSL pane lists the distribution through the \\wsl.localhost\ share
Windows serves it on, so what you see is the distribution's own filesystem and
not the Windows one — and ~ is resolved by asking the distribution once, never
by typing pwd into your pane.
Two things it will refuse, with a sentence rather than in silence:
- A pane that is not at a prompt. A
cdsent whilevi,less, or an agent has the terminal goes to that program instead. Where a shell reports the OSC 133 prompt marks — the same integration the command history uses — this is known exactly; otherwise the shape of the last line is the best available answer. - A directory name it cannot quote for that shell. A name containing a line break, or a double quote in a Command Prompt pane, is not typed at all rather than approximated.
The browser navigates; it does not move files. The SFTP panel is the surface for that.
Settings open from the gear at the bottom of the left rail, the avatar in the
title bar, or Ctrl+Shift+,. Changes apply immediately and are remembered
between launches; each page can be reset on its own.
- Appearance — theme, terminal font, size, line height and letter spacing, with a live preview. Panes restyle in place and keep their scrollback.
- Terminal — scrollback lines, cursor style and blink, and copy-on-select.
- Sessions — default shell and WSL distribution for new terminals, the layout to start in, and whether the sidebar starts collapsed.
- AI & voice — whether AI suggestions need approval before running, and whether a transcript is typed straight into the pane or shown for review first. Neither option presses Enter for you. Also the phrase the pane's proceed button sends, described below.
- Speech recognition — engine, model and tuning, described below.
Each pane's title bar carries a tick beside the microphone. Clicking it sends
OK, proceed and presses Enter — for the common case of an AI coding agent
pausing to ask whether it should carry on. The phrase is editable under
Settings ▸ AI & voice, so an agent that responds better to different wording
can have it.
This is the one control that presses Enter for you; voice transcripts and AI suggestions deliberately do not. It sends to the pane it belongs to, so a pane sitting at a shell prompt rather than in an agent will simply try to run the phrase as a command.
Two engines are available.
Built-in runs Whisper as ONNX inside the app through Transformers.js, with nothing else to install. Choose the model (tiny, base or small; English-only or multilingual), the weight precision, and whether to compute on CPU (WASM) or GPU (WebGPU) — WebGPU falls back to WASM when it is unavailable. The panel shows the download for the chosen combination, from about 41 MB for tiny at q8 to about 968 MB for small at full precision; models are cached after first use. Multilingual models add language and transcribe/translate options, which English-only checkpoints reject and so do not show.
Server posts the recorded audio as a WAV file to a transcription server,
using the OpenAI /v1/audio/transcriptions shape that whisper.cpp's server, LM
Studio, faster-whisper-server and similar tools speak. This is the way to use a
model the built-in engine cannot load — a GGUF build such as
unslothai/Qwen3-ASR-0.6B-GGUF needs a llama.cpp-family runtime, so something
else has to host it. Any http:// or https:// address works: loopback, a
machine on the LAN such as http://10.0.10.46:8888/v1/audio/transcriptions, or
a hosted service. The field says whether the address it holds is on this machine
or not, because a remote one means recorded speech leaves it.
Servers that want authentication take an API key in the same panel, sent as an
Authorization: Bearer header. The key is not kept in the settings file: the
main process stores it encrypted through the operating system's own secret store
— DPAPI on Windows, Keychain on macOS, libsecret or kwallet on Linux — and hands
it to the renderer only for the request being made. On a system with no keyring
available, saving is refused rather than written in the clear, and the key can be
held for the session instead. Leave it empty for a local server that wants none.
Both engines share the maximum utterance length and the silence threshold, and the Try it button on that page records a phrase and shows the transcript, the recording level and how long transcription took, without typing into a terminal. It transcribes even below the silence threshold and says so, which is how the threshold gets tuned for a particular microphone.
The simplest way to try the published package is through npx:
npx --package=electron@43 --package=zerogterm zerogtermThe npm launcher needs Electron alongside ZeroG Terminal; the command above installs both into the temporary npx environment. To use the launcher repeatedly without downloading on each invocation, install it globally:
npm install --global electron@43 zerogterm
zerogtermOpen a local shell in a specific directory (ZeroG FileMan uses this):
zerogterm --cwd "C:\Users\you\project"
# or: ZEROG_OPEN_CWD="C:\Users\you\project" zerogtermA second launch while ZeroG is already running focuses the existing window and opens another pane there (single-instance).
ZeroG Terminal can expose the running desktop instance as a local MCP server. This lets an MCP-capable AI agent inspect workspaces and sessions, restore or create project workspaces, open SSH sessions, and request commands for your approval.
-
Start ZeroG Terminal with MCP enabled:
ZEROG_MCP_ENABLED=1 npx --package=electron@43 --package=zerogterm zerogterm
On Windows PowerShell, use:
$env:ZEROG_MCP_ENABLED = "1" npx --package=electron@43 --package=zerogterm zerogterm
-
Open Settings → AI & voice → MCP in ZeroG Terminal. Copy the displayed endpoint and bearer token into your MCP client. The default endpoint is
http://127.0.0.1:60056/mcp.MCP clients that support Streamable HTTP generally need the equivalent of:
{ "url": "http://127.0.0.1:60056/mcp", "headers": { "Authorization": "Bearer <token-from-zero-g-settings>" } }The exact configuration key differs between clients; use its Streamable HTTP or remote MCP server setting rather than a stdio command. The token belongs to the current ZeroG process, is held only in memory, and must be copied again after ZeroG restarts. Never put it in source control, prompts, issue reports, or shared configuration.
-
Ask the agent to call
zerog_get_connection_status, thenzerog_acquire_controlwith a client ID and the capabilities it needs. Read capabilities require no write access; workspace changes and command requests require the corresponding lease capabilities. The lease is single-client and expires unless renewed.
Every command request is shown in ZeroG Terminal for exact user approval before anything is written to a shell. Rejection, cancellation, takeover, credential prompts, uncertain prompts, and unsupported command shapes fail closed. MCP does not receive terminal output, credentials, or keystroke-level control. See the full MCP protocol and security guide for tools, limits, and troubleshooting.
npm install
npm run typecheck
npm test
npm run build
npm startNode.js is the only prerequisite every host needs. The scripts above run on Linux
and Windows; npm runs them through cmd.exe on Windows, so no POSIX shell is
required whichever shell you start them from.
node-pty provides terminal I/O on every platform. It ships prebuilt binaries
for Windows and macOS, so a C/C++ toolchain is a Linux requirement rather than a
general one — there, npm install compiles it.
Download the latest release for your platform from GitHub Releases. Review the release notes before installing; SSH, WSL, PowerShell, Git Bash, and any remote screen sessions are provided by the host system rather than bundled with the app.
The Windows x64 installer adds ZeroG Terminal to the current user's applications. The portable .exe can be run without installation.
These binaries are unsigned, so Windows SmartScreen may show a warning on first launch. If you downloaded the release from the official repository, choose More info → Run anyway.
The Linux release provides an x64 AppImage and Debian package built for Ubuntu 22.04 or newer. Make the AppImage executable after downloading it, then launch it from the same directory. If the AppImage reports that FUSE2 is unavailable, install libfuse2 on Ubuntu 22.04 or libfuse2t64 on Ubuntu 24.04 and newer. The .deb package can be installed through Ubuntu's software tools or APT and does not require FUSE2.
The host still needs a C/C++ toolchain when developing from source:
sudo dnf install screen make gcc-c++ python3make, gcc-c++ and python3 build node-pty. screen is optional but worth
having: with it, local sessions are persistent and rediscovered after relaunch.
Without it, ZeroG falls back to a direct PTY on the chosen shell and labels the
session process-only; that shell is lost when the application exits.
A few seconds after launch, ZeroG asks GitHub whether a newer release exists, and shows a popup if one does. Nothing is downloaded until you choose to:
- Windows installer and AppImage: Download & install fetches the update (checked against the release's
latest.ymlhash), then Restart & install closes every pane and restarts into the new version. A downloaded update also installs the next time you quit. - Portable
.exeand.deb: these cannot replace themselves, so the popup offers Open download page instead.
Later hides the popup until the next launch, and Skip this version stops it for that version. Turn the startup check off, check by hand, or stop skipping a version in Settings → Updates. Development and npm-installed copies do not check.
No tooling beyond Node.js. The new-terminal dialog offers the shells it finds on PATH — Windows PowerShell, PowerShell 7, Command Prompt, WSL (with a distribution picker), and Git Bash where Git for Windows is installed.
screen does not exist on Windows, so local sessions are always process-only and
do not survive app exit. Remote SSH sessions are unaffected by the local shell
backend, and a remote host with screen still gives persistent sessions there.
The pinned node-pty dependency receives a small ConPTY teardown patch during
install and build (scripts/patch-node-pty.mjs). Review that patch before changing
its version. Run node scripts/pty-shutdown-smoke.cjs on Windows to check forced
and graceful shutdown with a real child process. Release evaluation evidence and
remaining sign-off items are in the release check report.
Windows source packaging rebuilds native dependencies. Install Visual Studio 2022
Build Tools with the C++ desktop workload by importing .vsconfig,
then run powershell -ExecutionPolicy Bypass -File scripts/ensure-windows-build-tools.ps1.
The helper checks the selected MSVC toolset and installs its missing Spectre libraries
through Visual Studio Installer if needed; Windows may request administrator approval.
Then run npm run package:win. The normal native rebuild and installed application
have been verified; compiler protections remain enabled.
Linux packaging uses /opt/ZeroGTerminal to avoid Electron's sandbox launcher issue
with spaces in installation paths; the application menu still says ZeroG Terminal.
The Debian package declares its audio and graphics runtime libraries explicitly.
The Electron downloader has a scoped global-agent override to remove the vulnerable
sprintf-js dependency chain. node scripts/check-downloader-proxy.mjs verifies proxy
routing and NO_PROXY bypass using local fixtures. CI runs this check and the full audit.
The current main branch has the following local verification coverage:
npm run typecheck: passes.npm test: 64 test files; 1,169 passed and 3 platform skips on Windows, all 1,172 passed on Linux withZEROG_LIVE_SCREEN=1.npm run build: passes and writesdist/mainplusdist/renderer.npm audit --audit-level=moderate: zero advisories in the verified dependency tree.
ZeroG Terminal is released under the MIT License. See LICENSE.
See CONTRIBUTING.md for development, testing, and pull-request guidance. Please report security vulnerabilities privately through GitHub; see SECURITY.md.