# Terax packs a terminal, editor, git graph, and an AI agent into an 8 MB binary > Source: > Published: 2026-07-26 14:24:04+00:00 Terax fits a terminal, a code editor, a git client with a real commit graph, a web preview pane, and an agentic AI panel into a binary that takes about 7 to 8 MB on disk. Most tools that pack that much surface area ship far heavier. Terax gets there by building on Tauri 2 and Rust instead of a bundled browser runtime, and the result is a dev workspace you install without an account and run without a single telemetry ping. The repo, crynta/terax-ai, calls itself an ADE: an AI-native dev environment. The terminal is the core. It runs on xterm.js with a WebGL renderer and a native PTY backend through `portable-pty` , so zsh, bash, pwsh, fish, and cmd all behave as real shells. On Windows, each tab can target a different workspace environment, either Local or any installed WSL distro, and the README is explicit that WSL is a first-class environment, not a wrapped subprocess. You get split panels, inline search, link detection, and true color. The editor is CodeMirror 6 with Vim mode and ten bundled themes, from Tokyo Night to Gruvbox to Xcode Dark. Source control is a full panel: stage and unstage individual hunks, commit with Ctrl+Enter, push with upstream awareness, and a history pane that draws an actual commit graph with lane rendering for merges and branches. The file explorer does fuzzy search, inline rename, and a Catppuccin icon theme. A web preview pane auto-detects local dev servers and opens them in a tab. None of this is a plugin you add later. It all lives in the base binary. What separates Terax from a plain terminal is the AI, and the choice worth noting is where the intelligence comes from. Terax is bring-your-own-key across a long list of providers: OpenAI, Anthropic, Google Gemini, Groq, xAI, Cerebras, OpenRouter, DeepSeek, Mistral, and any OpenAI-compatible endpoint. If you would rather not send code to anyone, it also talks to LM Studio, MLX, and Ollama for fully local inference. Inline autocomplete works against local models too. The agent side is more than a chat box. It plans multi-step work and confirms before acting, spawns sub-agents, and keeps project memory in a `TERAX.md` file. It carries the tool set you would expect from a coding agent: file read, write, edit, multi-edit, grep, and glob, plus bash execution behind approval gating and background processes. The composer pulls in snippets with `#handle` , files with `@path` , and supports slash commands and voice input. You can define custom agents, each with its own system prompt and a restricted tool subset. When the agent edits code, the diffs land in the editor and you accept or reject them hunk by hunk. Credentials are handled the way a local-first tool should handle them. API keys go into the OS keychain through `keyring` , and the README states they never touch disk or localStorage. Terax is early. It is not code-signed yet, so Windows shows a protective warning on first launch, and the open issue count is not small. But the positioning is consistent: a terminal-first workspace that treats the AI agent as a native part of the app rather than an extension, keeps everything on your machine unless you point it outward, and stays small enough that the download barely registers. Installers are on the Releases page and the app auto-updates from there. Arch users get `terax-bin` on the AUR, and Nix users get an official flake. The tech choices back the pitch: Tauri 2, Rust, `portable-pty` , React 19, TypeScript, xterm.js, CodeMirror 6, and the Vercel AI SDK v6. If you have wanted an agentic coding setup without a heavyweight footprint, a required login, or telemetry, crynta/terax-ai is worth a look. **GitHub:** [https://github.com/crynta/terax-ai](https://github.com/crynta/terax-ai) *Curated by Agent Palisade — practical AI for small and mid-sized businesses.*