Warp-style command blocks and a multi-step AI agent, without the bulk. Local models run in-process with Metal acceleration — no Ollama, no daemon to babysit. Cloud models sit behind the same interface.
Interface mockup of VTerminal's default theme — command blocks with exit-code badges, and the agent's approval gate showing both classification axes.
Why VTerminal
Modern AI terminals tend to be Electron apps that phone home for every completion. VTerminal is a native Tauri app that treats your shell, your machine and your consent as the things that must not be replaced.
Instead of a reimplemented shell
A real zsh -il login shell over a PTY. vim,
htop, tmux and ssh all behave exactly as they do
in Terminal.app, because they are running exactly as they do
in Terminal.app.
Instead of cloud round trips
The default model is a 9B GGUF running inside the app process via llama.cpp, Metal-accelerated. Nothing leaves the machine unless you pick a cloud model — and there is no daemon to install, start or keep updated.
Instead of an agent with free rein
Every command a model writes passes an approval gate. Natural-language suggestions are inserted into your prompt, never executed. Restored tabs never reconnect on their own.
The Terminal
Shell integration marks every command as a block, so the app knows what you ran, where, and how it ended — which is also exactly the context the model needs.
portable-pty spawning zsh -il, rendered
by xterm.js 6 with a WebGL renderer. Full TUI support, and saved
SSH hosts with one-click connect.
Exit-code badges, copy the command or just its output, re-run, or attach a block to the AI as context. Positions come from live terminal markers, so they never drift.
The reader thread blocks above a watermark and the UI
acknowledges every 256 KB, so cat-ing a
gigabyte file won't balloon memory.
Tabs with ⌘1…9, in-terminal search, persistent command history, and a ⌘K palette for actions and model switching.
Nested sessions are detected by command shape. While you're on a remote host, local paths and git branch are withheld from the model rather than quietly misreported.
Veviad Developer UI (default), Veviad UI, Midnight, Nord, Solarized Dark and Light — each with a matched terminal ANSI palette, not just recoloured chrome.
The AI
The same Provider trait covers in-process llama.cpp,
Anthropic, OpenAI, Mistral and any OpenAI-compatible server you
run yourself — so every feature below works the same whichever you
pick.
Describe the goal — or type # at an empty prompt —
and get a command inserted into your prompt, ready to read before
you press enter.
One click on a failed block streams a diagnosis of what went wrong, plus a corrected command you can inspect and run.
A chat panel with your blocks, their output and your files as context. Resizable, and it keeps its proportion when you resize the window.
Multi-step runs that propose commands, execute them in your visible terminal, and read the real output before deciding what to do next.
off → low → medium → high → max, per model, showing
only the rungs a model actually accepts — because a wrong value
is a 400, not a downgrade.
Drag, paste or pick. An optional on-device vision sidecar transcribes screenshots, so even a non-vision chat model can work with them.
Safety
This matters more in a terminal than anywhere else, so it is worth being precise about what is enforced and what is not.
Every proposed command is checked for whether it is read-only and whether it reaches the network — independently, since a fetch that writes no file still pulls unreviewed content into a loop that proposes shell commands. Both are shown on the card.
The permission mode is per session, never persisted, and never inherited by a new tab. Auto-run of reads requires read-only and no network — unknown commands fail closed and cost you one click.
Approved commands run in the tab you're looking at, not a hidden
subprocess — so you see exactly what ran, and it runs wherever
that tab is, including over ssh.
With web access off, the model's fetch tooling is never sent in the first place, and network-shaped commands are refused before an approval card is even drawn.
A command you rewrite before approving is treated as your own text on your own gesture, not as model output — deliberately not re-classified.
Restored sessions never replay anything. An ssh tab
you had open offers Reconnect; it does not reconnect
itself.
The command classifier cannot see through a script the agent
wrote in an earlier step, an alias in your dotfiles, a
base64-decoded one-liner, or python -c. It is
documented as best-effort in the app itself and should not be
relied on as a security boundary. What actually enforces "no
internet" for a capable model is the withheld tool, not the
string match.
Models
The catalog is a curated allowlist, so every model carries the reasoning-effort rungs it actually accepts and the RAM it actually needs — no runtime surprises, no 400s from a guessed parameter.
GGUF weights running inside the app via llama.cpp with Metal acceleration. Downloads are resumable and range-checked.
Qwen3.5 9B is the default, sized to run on a 32 GB M1 Pro.
Frontier models behind the same trait, with per-model effort mapping onto each vendor's own parameter.
Keys are stored by the Rust backend and never round-trip to the UI.
Point VTerminal at any OpenAI-compatible server, press Test, and pick which of the served models to expose.
<think> traces split out
Hosts are only contacted behind an explicit Test — never probed in the background.
A second local model loaded beside the chat model, purely to transcribe the images you attach.
Transcribed text is fenced and labelled as data — a screenshot is attacker-controllable by construction.
Switching models is one keystroke from the ⌘K palette, and every reply keeps the name of the model that wrote it.
Quick Start
VTerminal is source-first while it's pre-1.0. Clone it, build it, and you have the terminal plus cloud AI immediately.
rust-toolchain.tomlxcode-select --installbrew install cmake — only for the
local-llm feature
git clone https://github.com/Veviad/VTerminal.git
npm install
The terminal plus cloud AI, with a fast build that skips the local inference engine entirely.
npm run tauri dev
Then open Settings → Models, download Qwen3.5 9B (~5.3 GB) and press Load.
npm run tauri dev -- --features local-llm
The first local-llm build compiles llama.cpp and
its Metal kernels — budget 10–30 minutes. Incremental builds
afterwards are normal speed.
Keyboard
Open Source
Issues and pull requests are welcome — please open an issue before starting substantial work so the approach can be agreed first. If you intend to contribute regularly, get in touch: a licensing agreement may be needed to keep future relicensing possible.