Terax

Operator's take

The 7 MB figure is the first thing you will read about Terax, and it is the least useful reason to install it. A small binary tells you the app will open fast and stay out of your RAM. It tells you nothing about whether the thing inside is worth your working day. What is inside is a terminal, a CodeMirror editor, a file explorer, a git panel with a real commit graph, a web preview that finds your dev server on its own, and an AI agent that reads the repo and proposes diffs you accept hunk by hunk. That is a genuinely full workspace in a download smaller than a phone photo, and the reason it fits is Tauri: the app renders through the operating system's own webview instead of shipping its own copy of Chromium the way an Electron app does.

If you came here from a comparison of parallel-agent tools, the answer is no. Terax will spawn Claude Code, Codex, Gemini CLI, Pi, OpenCode or Grok into terminal panes, and it will tile up to four of them at once. It will also supervise one: its own AI reads the spawned agent's output, decides whether the task is done, and sends a follow-up instruction if it is not, for a capped number of rounds. But every one of those panes opens on the same working directory. The word "worktree" does not appear anywhere in the source tree. Four agents in four panes editing one checkout is not a fleet, it is a merge conflict with extra steps. Tools like cmux and Conductor exist specifically because isolating each agent in its own git worktree is the hard part, and Terax has not attempted it.

The cost of a young project is the part the marketing page cannot tell you. Terax is four months old, is written almost entirely by one person, and has an open issue count roughly equal to everything it has ever closed. Version numbers are still in the 0.8 range and the site's own footer calls the build a preview. The terminal engine is being swapped out underneath it right now: the public README and the About page both say xterm.js, while the project's own architecture doc on the main branch says xterm has been removed in favour of a Ghostty-based renderer. Releases came roughly weekly through July and then stopped for six weeks while that rewrite landed, even though commits kept arriving daily. None of that is misconduct. It is what a fast-moving solo project looks like from the outside, and it is the risk you are taking on.

So the adoption question is narrow. If your terminal, your editor and one AI agent living in one fast window would genuinely save you switching all day, and you can tolerate a preview-grade app you may have to file bugs against, install it: it is free, it is Apache-2.0, it takes nothing from you, and you can keep your existing terminal open beside it. If you are trying to run several coding agents in parallel without them treading on each other, Terax is not the tool that solves that, and no future release note can be assumed to change it.

What it's good at

• A real workspace in a tiny download — terminal, CodeMirror 6 editor, file explorer, source control and web preview in one app, at roughly 4–7 MB for the macOS, Windows and Linux package installers.
• Fast cold start — the site claims about 300 ms from launch to usable, on the strength of the Tauri and Rust stack rather than a bundled browser runtime.
• Bring your own model, or no cloud at all — direct keys for OpenAI, Anthropic, Google, Groq, xAI, Cerebras, OpenRouter, DeepSeek and Mistral, any OpenAI-compatible endpoint, or fully local inference through LM Studio, MLX or Ollama. Keys go to the OS keychain, not to disk.
• Reviewable AI edits — the agent's changes land as inline diffs you accept or reject hunk by hunk before they touch the file, and shell commands are approval-gated.
• Launches the CLI agents you already pay for — Claude Code, Codex, Gemini, Pi, OpenCode and Grok each get a one-click launcher into a terminal pane, with an optional supervision loop where Terax's own AI reads the output and sends follow-ups for up to three rounds.
• Nothing to sign up for — Apache-2.0, no account, no telemetry, no paid tier, and a stated position that there is no upsell coming.
• Genuinely cross-platform — macOS on Apple Silicon and Intel, Windows with WSL treated as a first-class workspace, and Linux via .deb, .rpm, AppImage, AUR and a Nix flake.

What it's not

• Not a parallel-agent orchestrator — it tiles up to four agent panes, but they share one working directory and there is no git worktree isolation anywhere in the codebase. For running a fleet safely, look at cmux, Conductor or Herdr instead.
• Not a mature product — the repo is four months old, sits at v0.8.6, is labelled a preview on the vendor's own site, and carries roughly as many open issues as it has ever closed.
• Not a team with a support contract behind it — one maintainer has written over 90% of the commits; the community contributions are real but small. If he stops, it stops.
• Not a full IDE — the maintainer says so directly: no debugger UI, no refactoring tools, and language-server support is opt-in rather than automatic.
• Not actually 7 MB on every platform — the Linux AppImage is about 90 MB. The small number describes the .dmg, .deb, .rpm and .exe, not every build.
• Not something with an independently verified speed claim — the ~300 ms cold start and the sub-millisecond keystroke latency are the vendor's own numbers, published without a methodology.
• Not for someone who wants their agent in the cloud — there is no hosted service, no remote session, and no way to check on a running agent from a phone. It is a local desktop app and nothing else.