Gemini CLI by Melhem Shoker

Unofficial. Not affiliated with, endorsed by, or sponsored by Google. "Gemini" is a trademark of Google LLC. This is an independent front-end for the open-source Gemini CLI.

A Gemini chat panel in the VS Code sidebar. It runs the Gemini CLI as a persistent agent over the Agent Client Protocol, so it streams reasoning, asks before touching your files, and answers in about a third of the time a fresh-process approach takes.

Setup

Nothing to configure by hand. Open the panel and it checks three things:


Node.js	needed to run the CLI
Gemini CLI	one click installs it via npm
API key	one click, free from AI Studio

The key goes into VS Code Secret Storage - encrypted by the OS keychain, never written to a file in your project, never committed, never bundled. It is injected into the agent process as an environment variable at spawn time and nowhere else. An existing ~/.gemini/.env or a GEMINI_API_KEY environment variable is picked up automatically if you already have one.

Why it is fast

The obvious design spawns gemini -p "..." per message and pays process startup every time. This runs one gemini --acp process for the workspace instead:

	spawn per message	persistent agent
startup	~1.9s every message	~1.9s once
warm turn	~7.2s	~2.5s

Conversation context lives in the agent session, so follow-ups cost nothing extra to send.

What it does

Streams reasoning - a collapsible Thinking section, separate from the answer
Approval modes - Default / Auto Edit / YOLO / Plan (read-only), switchable live from the header without restarting the session
Inline permission prompts - in Default mode the agent asks before each tool, with Allow/Reject buttons and the exact file it wants
Slash commands - type / for the agent's own commands (memory, tools, and more)
@file context - type @ to fuzzy-search workspace files
Tool calls - live rows with a kind glyph, click to open the file
Editor-aware - the agent reads your unsaved buffer, not the version on disk
Custom tools - your own one-click prompt buttons
Transcript survives window reloads

Custom tools

geminiMS.customTools renders a button row. Four placeholders:

Placeholder	Expands to
`${selection}`	selected text, or the whole file if nothing is selected
`${file}`	absolute path of the active file
`${fileName}`	basename of the active file
`${language}`	VS Code language id

"geminiMS.customTools": [
  { "label": "Arabic", "prompt": "Explain in Lebanese Arabic:\n\n${selection}" },
  { "label": "SQL",    "prompt": "Optimize this query:\n\n${selection}" }
]

Ships with Explain / Fix / Tests / Refactor / Review / Docs. The row refreshes on save.

Settings

Setting	Default	Purpose
`geminiMS.approvalMode`	`default`	starting mode for new sessions
`geminiMS.includeActiveFile`	`true`	prepend active file path + selected range
`geminiMS.eagerStart`	`true`	warm the agent when the panel opens
`geminiMS.customTools`	6 presets	your prompt buttons
`geminiMS.nodePath`	`""`	override the `node` executable
`geminiMS.geminiJsPath`	`""`	override CLI auto-detection

Keyboard

Key
`Ctrl+Shift+G`	focus the chat
`Enter` / `Shift+Enter`	send / newline
`Esc`	cancel the current turn
`Up` `Down` `Tab`	navigate the `/` and `@` popups

Commands

Gemini: Ask... / New Session / Cancel / Set API Key / Clear API Key / Install Gemini CLI / Settings

Releasing

Add a repository secret VSCE_PAT (an Azure DevOps token scoped to Marketplace > Manage), then:

npm version patch
git push --follow-tags

.github/workflows/publish.yml syntax-checks, verifies the tag matches package.json, packages, publishes, and attaches the .vsix to a GitHub Release.

Requirements

Node 20+ / @google/gemini-cli / a Gemini API key. The panel installs the last two for you.

MIT (c) Melhem Shoker