Pack de Démarrage Extensions Gemini CLI

What's in this pack

Gemini CLI hit GA, the official extension list grew to ~10 entries, and most tutorials walk you through the same four (Stitch, Workspace, Code Review, Postgres). Fine for a demo. Useless if you actually work in the Google stack day to day.

This pack is for the developer who chose Gemini CLI deliberately — because of the 1M-token context window, the free 60 requests/min tier for Gemini users, the native multimodal input (screenshots, PDFs, video frames feed straight in), and the deep Google ecosystem hooks (BigQuery, Firebase, GCP) that Claude Code and Cursor have to bolt on as third-party MCPs. You want the picks that lean into those strengths instead of treating Gemini CLI as a Claude Code clone.

Ten entries: the CLI itself, the five official extensions that get under-discussed (the BigQuery / gcloud / Genkit / Firebase / Flutter quartet plus Cloud Assist), and four high-leverage MCPs that extend Gemini's native vision and Google-app reach. By install #10 you have a coding agent that drafts SQL against your real warehouse, deploys Firebase functions, designs Flutter screens via Stitch, reads Gmail + Drive, and queries NotebookLM corpora — none of which is bolt-on theater. All Google-native.

Install in this order

1. Gemini CLI — install the base agent first. npm i -g @google/gemini-cli, sign in with a Google account, you get 60 req/min free (Gemini API user) + 1M-token context. Don't skip the tier check: the free quota is the reason this stack is even viable for solo devs.
2. Gemini CLI Extension: gcloud — the foundation. Once gcloud is bound to natural-language commands, every later GCP-touching extension (BigQuery, Firebase, Cloud Assist) inherits sane authentication and project scoping. Install gcloud second; everything else depends on gcloud auth login being healthy.
3. Gemini CLI Extension: BigQuery — the data-analytics killer feature. Gemini's 1M context means you can paste a 500-table schema, ask "why did Q4 revenue dip," and get a real SQL draft. This is the use case Claude Code can't match without a third-party MCP and 4 hours of setup.
4. Gemini CLI Extension: Firebase — deploy functions, manage Firestore rules, configure auth — all from chat. If you build mobile / web apps and your backend is Firebase, this collapses a firebase deploy workflow into one prompt.
5. Gemini CLI Extension: Flutter — pair with Firebase (#4) for the canonical Google mobile stack. Widget generation, state management scaffolding, platform-specific code. Multimodal angle: paste a screenshot of the UI you want, get the widget tree.
6. Gemini CLI Extension: Genkit — the AI-app framework. When you stop using Gemini and start embedding it (RAG flows, structured outputs, retrievers), Genkit is the SDK. Extension makes flow scaffolding a 30-second job.
7. Gemini Cloud Assist MCP — Private Preview MCP server that surfaces Google Cloud Ops APIs (logs, traces, IAM diagnostics) to any MCP client. Pairs with gcloud (#2): one talks to GCP, this one talks about GCP — debugging a 502 in production means asking Gemini to read the actual Cloud Logging stream.
8. Google Stitch MCP — the multimodal showcase. Paste a Figma URL, an old screenshot, or a hand-drawn wireframe — Stitch generates a real UI. Wired through MCP, Gemini can iterate ("make the CTA blue, add a search bar") and the design system stays consistent. This is where Gemini's vision tokens earn their keep.
9. Google Workspace MCP — official Google CLI with a built-in MCP server for Gmail, Drive, Calendar, Docs, Sheets and 10+ services. Once installed, Gemini reads your inbox, drafts replies, creates calendar holds, and edits Docs in place. The mundane productivity layer that turns CLI from "coder" into "assistant."
10. NotebookLM MCP — the nlm CLI plus MCP server lets Gemini operate NotebookLM notebooks: upload sources, generate audio overviews, query against a private RAG corpus. Pair with Workspace MCP and you have a complete "read everything I work with, then reason" loop — the long-context use case Gemini was literally built for.

How they fit together

Gemini CLI (#1)
   │
   └─ 1M context · multimodal · free 60 rpm
         │
gcloud extension (#2)  ←─ foundation: auth + project scope
   │
   ├─ BigQuery (#3) ────────── data + SQL
   ├─ Firebase (#4)  ─┐
   │                  ├─ mobile / web app stack
   ├─ Flutter (#5)  ──┘       (screenshot → widget tree)
   │
   └─ Genkit (#6) ─────────── build AI features yourself

Google ecosystem MCPs (drop-in, not extensions):
   ├─ Cloud Assist MCP (#7) ── debug GCP prod (logs / traces)
   ├─ Stitch MCP (#8)  ─────── multimodal design (image → UI)
   ├─ Workspace MCP (#9) ───── Gmail / Drive / Calendar / Docs
   └─ NotebookLM MCP (#10) ── long-context RAG over your sources

The load-bearing trio is CLI (#1) + gcloud (#2) + one MCP that fits your workflow. If you're a data person, that MCP is Cloud Assist (#7). If you're a designer/PM, it's Stitch (#8). If you're an ops generalist, it's Workspace (#9). Don't install all four MCPs day one — pick the one matching what you do, add others as you hit the gap.

Tradeoffs you'll hit

• Gemini CLI vs Claude Code vs Cursor — Different bets. Claude Code is the strongest agent (best at multi-step reasoning, sub-agents, hooks). Cursor is the strongest editor (in-IDE diff UI, tab-complete). Gemini CLI is the strongest Google-native multimodal tool (1M context, free tier, BigQuery / Firebase / Stitch first-class). Most pros run two: Claude Code or Cursor for daily code, Gemini CLI when they need to dump a PDF or screenshot in.
• Official extensions vs MCPs — Extensions (#2-#6) are first-party, ship signed, no setup. MCPs (#7-#10) are richer but require a JSON config and sometimes a separate auth flow. Lean on extensions where they exist; reach for MCP when no extension covers the target.
• Stitch MCP vs Stitch extension — The existing official Stitch extension is fine for inside-CLI use. The MCP version is the same capability exposed over MCP so Cursor and Claude Code can call it too. Install whichever your daily editor speaks; install both if you switch.
• Workspace MCP vs Gmail API directly — Workspace MCP is one config away from "AI reads my inbox." Raw Gmail API is two hours of OAuth scaffolding for the same thing. If you ever found yourself writing a Google Apps Script to forward emails, the MCP route is strictly better.
• Free tier reality check — 60 rpm sounds generous but multimodal calls (image / video frames) burn it faster. If you'll be pasting screenshots all day, budget for a paid Gemini API key before week 2.

Common pitfalls

• Skipping gcloud (#2) because "I don't use GCP" — Even if your prod is on AWS, the gcloud extension is what wires gcloud auth to the BigQuery / Firebase / Cloud Assist trio. Install it; ignore the GCP-only commands.
• Installing every extension — There are ~10 official extensions. You don't need all of them. Pick the ones touching your stack. Extra extensions inflate the prompt context Gemini has to scan to route a request, which makes everything slightly slower.
• Pasting a 500-page PDF on a free tier — Gemini will accept it, parse it, and you'll watch your daily quota evaporate in one call. Chunk first, or upgrade to paid.
• Wiring Workspace MCP into the wrong Google account — Easy to OAuth into your personal account when your work data is in @company.com. Double-check the consent screen.
• Treating Stitch MCP output as ready-to-ship — It's a great first draft. Hand it to a designer before pushing to prod. The aesthetic is competent-default, not on-brand.

Why this complements the other Gemini pack

TokRepo already has a Gemini CLI Extensions pack covering the 10 official extensions Google ships (Stitch, Workspace, Code Review, Postgres, Vertex AI, and others). That one is a completionist index — install everything, see what sticks.

This pack is the opposite curation: a small, opinionated, multimodal-first short list mixing the CLI itself, the five less-hyped official extensions, and four MCPs that extend Gemini into Google Workspace, NotebookLM, Stitch, and Cloud Ops. If the official pack is the spec sheet, this is the playbook.