CLI Launcher for Claude, Codex, Grok, Kiro & Antigravity

Run Claude Code, Codex, Grok, Kiro & Antigravity — five AI coding CLIs — inside VSCode tabs, read the conversation as live markdown, manage sessions in a tree, and auto-sync your workspace with git.

A VSCode / Antigravity extension for the Claude Code, Codex, Grok, Kiro, and Antigravity CLIs.

한국어 README

I use Claude Code every day because I like it. Then Codex, Grok, Kiro, and Antigravity each found their way into my workflow too. AI coding CLIs are great in a plain terminal — but the more of them I kept open, the clearer it got that one terminal isn't enough.

• Every agent needs its own terminal.
• Session history is scattered, and each CLI stores it differently (some jsonl, some SQLite).
• I keep hopping between windows to check whether a response finished. On long thinks I can't get anything else done while waiting.
• Every file path / URL / folder in a reply has to be retyped or copy-pasted by hand.
• Frequently-used commands (/init, prompt prefixes) get retyped every time.
• With the same workspace on two or more devices, I keep forgetting to git pull / git push.
• I burn through my usage (5-hour / 7-day) without noticing, then suddenly hit the wall.

I built cli-launcher to solve all of that. It started by wrapping just Claude Code; now it runs five CLIs the same way inside one panel.

This is what it looks like. A session runs inside a VSCode tab with an input box at the bottom. Tabs along the top run several sessions at once, and the HUD reports your usage.

What it does

In one line: run Claude Code, Codex, Grok, Kiro & Antigravity in VSCode tabs, read the conversation as clean markdown, manage sessions in a tree, and auto-handle your workspace's git push/pull.

A little more:

• The CLI as-is, with conveniences on top — every feature of each CLI stays 100% intact. The launcher only layers the friction-fixers on top; it never hides or replaces anything.
• Five agents, one panel — run Claude / Codex / Grok / Kiro / Antigravity in the same Webview tab. Each agent gets its own sessions view, and each tab wears its agent's color.
• Session tree + groups — give each session a title and bucket related ones into groups/folders. Re-open an archived session from the tree to pick up where you left off.
• Reader-Live — watch the running conversation as markdown — code blocks, quotes, tables all rendered — side by side with the xterm.js terminal in a split layout. (Antigravity excepted — see "The five agents".)
• Status + notifications — a tab and its panel border glow yellow while a response is in flight and turn green when it's done, plus a desktop notification so you notice from another window.
• Hand-off — pass one agent's conversation context to another. "Did this much with Claude, now continue with Codex."
• Repo Sync — auto commit + push on workspace changes; the device name is stamped into the commit message.
• Multi-account switching — save Claude logins as profiles and swap between them. (file-based on Windows/Linux, Keychain-based on macOS.)
• HUD — 5-hour / 7-day usage rate + next reset time.

The five agents

cli-launcher wraps five AI coding CLIs the same way. Each one stores sessions differently (file vs DB, who assigns the id), so the internals differ — but from the outside it's identical: open it in a tab, pick it from the tree, read it in the reader.

• Only Claude is enabled by default. Turn the rest on under Settings → Agent; whichever ones are actually installed then show up in the new-session picker and the sidebar sessions view.
• Tabs take their agent's color automatically (Auto theme). The terminal output colors are left alone — only the bottom input area's tone changes — so you can tell which agent you're in at a glance while output stays consistent. You can also pin one color everywhere.
• Only Antigravity has no reader. agy stores its conversations as protobuf-in-SQLite, which can't be unpacked into markdown, so Antigravity tabs open terminal-only. Hand-off to Antigravity works (the context is injected into the prompt), but handing off from it doesn't — there's no readable transcript to carry over.
• Each permission toggle is that agent's "don't ask, just run" mode. Use it only on trusted work; all of them are off by default.

Why I built it

First, session visibility. A CLI's TUI is fine, but scrolling up through a long conversation isn't. With a reader, the session file is read straight back as clean markdown — code highlighting, tables, links all intact.

Identifying the session itself follows the same idea. Each CLI tracks sessions by an internal id only, so after a while you can't tell which was which. cli-launcher lets you give each session a title and bucket related ones into groups/folders. From the left tree you can re-open an archived session and keep working. Skip the title and it auto-generates one from the first message, so a session is always identifiable.

You see the same session two ways at once: as markdown on top, as the real terminal below. The top pane isn't read-only — you can send a message from the input box at the very bottom. Drag the splitter in the middle to rebalance.

Second, response state has to be visible at a glance. When the model thinks hard it can take over a minute, and going to check whether it finished broke my flow every time. So I made the state visual — the tab and panel border glow yellow while running and turn green when done, with a desktop notification too. Working in another window, a glance at the color tells you it's done.

Third, multiple devices. I move between a work PC, a home desktop, and a Mac, and every time I'm unsure whether the other side pushed. Laziness at shutdown breeds anxiety at startup. Repo Sync removes the shutdown-side laziness, which dissolves the startup-side doubt too: workspace changes are debounced for 5 minutes and auto committed + pushed, a final commit fires on VSCode exit, and the device name is stamped into the message ([Mac] auto-sync: …).

Fourth, usage visibility. I wanted to see Claude Code's rate-limit info in real time. Pinned to the status bar, a glance is enough to pace myself.

Fifth, the CLI stays; the flow just gets smoother. cli-launcher's most important principle is to leave every feature of each CLI 100% intact and only layer the friction-fixers on top — never hiding or replacing anything.

Reaching into a reply to grab a file path by hand got old, so file/URL/folder mentions in the response text are automatically clickable — one click opens the file in the IDE, a URL in the browser, a folder in Finder. Frequently-used commands go on custom buttons that drop into the input on click. Paste an image and it's saved to a temp file with just the path (@/tmp/...png) dropped in; large text gets converted to a .txt the same way, so the input box never gets clogged.

Features

• Per-agent session views — a sidebar tree for each of Claude / Codex / Grok / Kiro / Antigravity. Groups, folders (nested up to 3 levels), drag-and-drop, rename, trash. Recent Sessions is split by period (Today, Yesterday, Previous 7 / 30 days, then by month), so every ungrouped session stays reachable.
• Reader-Live — the running conversation as markdown, split-pane or standalone, with a branded welcome screen for empty sessions. (Antigravity excepted.)
• Customizable reader sender names — set the name shown for "you" (global) and a per-agent AI name in Settings.
• Status-aware tabs — idle / running / done / error / needs-attention. Interactive prompts ([Y/n], menus) skip the 7-second threshold and fire needs-attention + a desktop notification immediately. A tab that needs you while you are on another one blinks its icon green ↔ gray; the title stays as it is, so the other tabs do not move.
• Agent themes — the input area's tone follows the agent (Auto); pin one tone if you prefer.
• Hand-off — pass conversation context to any other enabled agent.
• Session links — right-click a session (or hit 🔗 in the toolbar) to copy a link that reopens it. Paste it into a task tracker, a note, or a chat; clicking it opens the editor and resumes that exact session, in the window that owns its folder.
• Slack inbox — "⋯ › Send to Claude" on any Slack message asks What should Claude do? in a small Slack box, then opens a new Claude tab with your words and the message as its first prompt (sender, channel, time, permalink, attachments, then the message quoted). Leave the box empty to send just the message. Connects over Slack Socket Mode, so no public URL is needed. The Slack app needs Socket Mode on, an app-level token with connections:write, the users:read bot scope, and a message shortcut with Callback ID send_to_claude (Interactivity & Shortcuts). Then run Claude: Set up Slack Inbox… with the bot and app-level tokens (kept in the editor's secret storage). Only the member IDs you list can start a session; anyone else gets a short "not allowed" reply. The tab opens in the folder you pick, in the window that has it open. These settings are read from user settings only, so a repo's .vscode/settings.json cannot change who may use it.
• cli-launcher for agents — every tab gets a cli-launcher command on its PATH, so the agent inside (Claude Code, Codex, gjc, Kiro, Grok, Chief) can work with its own tab: cli-launcher link prints the resume link for its session (ask it to attach that to the todo it files, the note it writes, the PR it opens), cli-launcher rename "<title>" names the tab, cli-launcher tabs lists open tabs with their links, cli-launcher open <link> asks you to resume a session (you click Open), cli-launcher notify "<message>" flags the tab and shows a desktop notification. Claude Code tabs wire their Stop / Notification hooks to notify, so "done" and "needs input" come from Claude itself rather than from quiet terminal output (no false alarms during long tool calls). No setup or per-agent registration; it runs on the editor's own runtime and talks to this window over a local pipe. cli-launcher help lists the options. Turn it off with claudeCodeLauncher.agentCli.enabled.
• Multi-account switching — save / switch Claude logins as profiles. The active account shows in the left status bar; click for a QuickPick. Tokens are read/written from ~/.claude/.credentials.json on Windows/Linux and from the Keychain on macOS.
• Auto-links + context menu — click file/URL/folder mentions in a reply; right-click a selection for a menu.
• Find files anywhere — clicking a bare filename that lives outside the session's working directory falls back to the OS file index (Everything on Windows, Spotlight on macOS, locate on Linux) and opens it; one hit opens directly, several show a picker. One-time setup in Getting started.
• Custom buttons / smart paste / input history / task queue — input-panel conveniences.
• Slash autocomplete — the Claude Code built-in catalog plus your personal PKM/OMC catalog (via an override file).
• Repo Sync — auto commit + push, with ${workspaceFolder} substitution.
• HUD — 5-hour / 7-day usage + reset time + model name.
• Optional tmux/psmux backend — run the session inside a multiplexer in the same Webview tab, for external attach + multi-machine workflows.
• Optional OMC mode — with oh-my-claudecode installed, a CCG viewer + HUD light up.

Getting started

Install — three ways:

1. Open VSX: search for CLI Launcher for Claude, Codex, Grok, Kiro & Antigravity in the Extensions view of Antigravity, or VSCode with the Open VSX gallery.
2. VSIX: download the latest cli-launcher-for-claude-<platform>-<version>.vsix from GitHub Releases and install via Extensions: Install from VSIX....
3. Build from source:

   git clone https://github.com/rockuen/cli-launcher-for-claude.git
   cd cli-launcher-for-claude
   npm install && npm run build && npm run package
   

At minimum you need claude on PATH (npm install -g @anthropic-ai/claude-code or the official standalone install). For Codex / Grok / Kiro / Antigravity, install each CLI and enable it under Settings → Agent.

Optional — "find files anywhere" lets you click a file link that lives outside the session's working directory and have it located via your OS file index. On by default; it degrades gracefully if the backend isn't installed.

• Windows — install Everything (keep it running) plus its command-line tool es.exe. Put es.exe in %LOCALAPPDATA%\Programs\everything-cli\ (auto-detected) or set claudeCodeLauncher.fileLocator.esPath.
• macOS — nothing to install; Spotlight's mdfind is built in.
• Linux — install plocate (or mlocate) and seed the db once: sudo apt install plocate && sudo updatedb.

Basic use:

1. Click the launcher icon in the editor title bar (or Cmd/Ctrl + Shift + ;), then pick the agent (Claude / Codex / Grok / Kiro / Antigravity).
2. Start working. To open another tab with the same agent, use the in-session +.
3. Re-open an archived session by clicking it in the per-agent Sessions tree (sidebar → CLI Launcher activity).
4. (Optional) Toggle the 👁 in the header for a split layout — reader on top, terminal below.
5. (Optional) Enable Repo Sync in the gear modal (⚙) and set your workspace path.

Repo Sync can be toggled in the in-panel ⚙ Settings modal, or set directly in your workspace .vscode/settings.json:

// .vscode/settings.json (workspace)
{
  "claudeCodeLauncher.repoSync.enabled": true,
  "claudeCodeLauncher.repoSync.path": "${workspaceFolder}",
  "claudeCodeLauncher.repoSync.deviceName": "Mac"   // prompted on first enable
}

It supports ${workspaceFolder} / ${userHome} substitution, so you never hard-code an absolute path per device. Commit .vscode/settings.json into the vault repo and it fits every device automatically.

Who it's for

• Anyone who uses Claude Code (or Codex / Grok / Kiro / Antigravity) daily and finds the bare TUI cramped.
• Anyone juggling several AI coding CLIs who wants to run them the same way in one panel.
• Anyone unifying their workflow inside VSCode/VSCodium.
• Anyone moving between devices on the same vault/repo.
• Anyone who wants their usage visible at a glance.

One honest note

• The HUD works best with OMC (oh-my-claudecode) installed, because it reads the usage cache OMC produces. Everything else (TUI / Reader / Repo Sync / multiple agents) works without OMC; only the HUD stays empty.
• Antigravity has no reader. Its conversations are protobuf-in-SQLite and can't be rendered as markdown, so it runs terminal-only and hand-off is receive-only.
• Multi-account switching is Claude-only for now. Tokens live in the Keychain on macOS and in a file elsewhere.

Closing

It's a small tool, but building it tidied up the flow I use every day. Even as the agents grew from one to five, the things that used to scatter all collect inside one panel.

I hope it helps people who lean on AI coding CLIs. Thoughts / issues welcome at GitHub Issues.

Full changelog: CHANGELOG.md.

License

MIT. Made by @rockuen.