# Line CLI > The fast terminal coding agent by Clatter Labs. Written in Rust, runs on SuperCosmos 1.1, with parallel sub-agents, MCP and native Arabic. Line CLI (`line`) is a coding agent that lives in your terminal. It reads and edits your project, runs commands with your approval, and splits big jobs across parallel sub-agents on your machine. ```bash curl -fsSL https://install.clatterlabs.xyz/line | sh # macOS / Linux line login # sign in with your Clatter Labs account line # start the agent in your project ``` macOS, Linux and Windows. Slash commands, flags and keys. Teach Line repeatable procedures with SKILL.md. Add tools, commands and hooks. ## Features - **Streaming TUI** with Markdown, code and table rendering, multiline input (Shift+Enter, Alt+Enter or Ctrl+J), scrolling with the mouse wheel or PgUp/PgDn (the view stays put while output streams), Esc to interrupt, Ctrl+C to exit. - **SuperCosmos 1.1**: **[SuperCosmos](/models/overview)** (efforts Flash, Max, Heavy, plus `auto` routing; 550B parameters, 55B active, 1M-token context). Switch effort with `/effort`. Line Code was retired on Nov 1, 2026; a saved Line Code choice moves to SuperCosmos automatically ([details](/models/line-code-retired)). - **Local tools** with approval prompts: read, write, edit and multi-edit files (diff preview), move and delete, list folders, glob, grep, run shell commands (in the background too, for dev servers and watchers), read-only `git`, `diagnostics` (your project's own type checker, compiler or linter), a **repo map** (a ranked outline of the codebase, built with tree-sitter), Jupyter notebook edits, web search and fetch, and a visible to-do plan. - **Parallel sub-agents**: the agent starts sub-agents on your machine to explore, research and review (read-only) or to do real work with full tools (general). Line locks files per agent, makes each agent read a file before editing it, and refuses an edit when the file changed since that agent read it, so parallel edits never overwrite each other. The **Agents panel** shows them live; `/agents` lists runs and their reports. - **MCP**: connect MCP servers (stdio or HTTP) in `~/.line/mcp.json` or your project's `.mcp.json`; their tools appear to the agent. `/mcp` shows them, `line mcp` lists them from the shell. Project servers start only after you trust them. - **Hooks**: run your own scripts on PreToolUse, PostToolUse, UserPromptSubmit, Stop and SessionStart (`/hooks`). - **Project instructions**: `LINE.md` (or `AGENTS.md`) in the project and `~/.line/LINE.md` are always followed; `/init` writes one. - **See every change**: `/diff` shows this session's changes (`/diff split` side by side), `/files` lists them by turn and agent; every file change is checkpointed, so `/undo` and `/rewind` take it back. - **Context and usage**: a live context meter in the status bar, `/context` breaks the window down, and auto-compact starts at about 80% of the model's real context window. `/usage` shows how much of your plan's limits you've used and when they reset, credits, and tokens and cost per model; `/cost` this session. - **Arabic, natively**: right-to-left text is shaped (joined letters), shown in the right order and right-aligned, mixed with English and code, in answers and in the input box. `/rtl` switches it (`terminal` if your terminal already handles bidi). - **Plan mode and permission modes**: Shift+Tab cycles ask → accept edits → auto → plan. - **@file mentions**, a live to-do panel, and a status bar with tokens, context and files changed. - **Busy models don't fail your turn**: Line waits and retries with backoff, and shows it. - **Skills and plugins** you control: nothing runs until you approve or enable it. - **Sessions** saved locally; `line -c` continues the last one in a folder. ## Safety defaults - Writes and shell commands ask first. **Shift+Tab** cycles the mode: ask, accept edits (file edits run, commands still ask), auto (nothing asks), plan (read-only); `line -y` starts in auto. Sub-agents follow the same mode; their approvals queue in the main window. - Files that usually hold secrets (`.env`, SSH and cloud credentials, private keys) **always** ask, even with auto-approve on. - Read-only commands that run without asking are limited to the project folder. - Project MCP servers and hooks run only after you trust them. - Commands Line runs never see Line's own sign-in token. ## Pricing Line is covered by your [plan](/billing/credits) (Free included), with usage valued at the SuperCosmos [rate](/billing/rates) of the effort each call ran on (see [Models](/models)). Every plan, Free included, includes every effort.