Gemini CLI — Module 22.6

How This Lesson Fits the Module & Volume

Claude Code showed a terminal agent on Anthropic. Gemini CLI is Google’s peer: an open-direction command-line coding agent that calls Gemini models (AI Studio, Gemini API, or Vertex AI). Cross-link Vol. 22.1 Google / DeepMind / Gemini / Vertex and Vol. 18 Gemini SDK—same substrate, different job (developer agent vs your FastAPI backend).

Gemini’s long context (Vol. 11 context length) is the catalog differentiator versus typical IDE Tab windows. Vol. 15 agent loops + HITL still apply. After this lecture the module leaves terminals and IDEs for prompt-to-app builders: Bolt.new.

Learning Objectives

By the end of this lesson, students should be able to:

Describe Gemini CLI as a terminal coding agent on Gemini / Vertex, not an IDE autocomplete plugin.
Explain when long context helps repo Q&A vs when retrieval + ignore files still win.
Authenticate via AI Studio API key vs Vertex IAM without mixing consumer and VPC data terms.
Compare Gemini CLI to Claude Code, Copilot CLI, and GUI IDEs qualitatively.
Apply HITL tool permissions and Vol. 20 privacy to shell/file tools.
Know when Google-estate teams should standardize here vs Anthropic or GitHub.

Definition

Gemini CLI is Google’s command-line interface for agentic coding (and related developer tasks) using Gemini models. It typically runs in a repository, uses tools (read/write, shell, search), and can attach large context or project instruction files (names such as GEMINI.md evolve—check current docs). It may be distributed as an open-source CLI that you point at AI Studio, the Gemini API, or Vertex. It is not a replacement for the Gemini SDK inside your product, and it is not VS Code Tab.

Agentic CLI vs Autocomplete vs Long Context

Mode	What Gemini CLI is	What it is not
Agent loop	Plan → tools → edit → test (Vol. 15)	Unsupervised prod deploy
Autocomplete	Optional via IDE plugins elsewhere (Gemini in VS Code/Android Studio, etc.)	This CLI’s headline job
Long context	Can ingest large specs / many files when the window allows	A reason to paste `.env` or ignore Vol. 11 lost-in-the-middle
Multimodal	Screenshots / diagrams into the CLI when supported	A substitute for Vol. 19 eval

Lean on long context

One-shot “read this RFC + these 8 files”
Migration notes that must stay consistent
When retrieval indexes are immature

Still retrieve / ignore

Monorepos larger than the useful window
Secrets, lockfiles, build artifacts
High-frequency Tab-like edits (use an IDE)

Vertex vs AI Studio

Vertex: IAM, VPC-SC, enterprise DPA
AI Studio keys: fast student labs
Do not mix prod customer data into Studio keys

Privacy and Repo Context

Same rules as Claude Code: the agent sees what it reads. Google data-use terms differ between consumer Gemini, AI Studio, and Vertex—Vol. 20 privacy means you pick the SKU before pointing the CLI at a private repo. Ignore secrets. Prefer project-scoped workdirs. HITL on shell. Long context is not a privacy feature; it can make over-sharing easier.

Catalog Snapshot (Qualitative)

Dimension	Gemini CLI	Claude Code	Cursor / Copilot
Pricing posture	Gemini API / AI Studio / Vertex usage (and any CLI-related quotas); possible free-tier experiments—live SKU sheet	Anthropic plan / API tokens	IDE seats + premium/agent meters
Strengths	Long context; Google Cloud alignment; open CLI; multimodal inputs; pairs with Vertex	Mature agent UX; Claude coding quality; `CLAUDE.md` culture	Ghost text + visual diff; Copilot multi-IDE
Weaknesses	Newer habit-share than Copilot/Cursor; IAM/setup cost on Vertex; not a full IDE	Not Google-estate native; no Tab	Weaker huge-context paste story; GitHub/VS Code gravity
Agentic vs autocomplete	Agentic CLI default	Agentic CLI default	Autocomplete + agents
Privacy / repo	Choose Studio vs Vertex terms deliberately; ignore + HITL tools	Anthropic retention SKU + permissions	Privacy Mode / content exclusion

CLI Pattern Sketch

Binary names, flags, and instruction filenames change. Verify against current Gemini CLI docs. Educational only; repos you maintain.

# GEMINI.md — project instructions (no secrets)
# Use Vertex in prod labs; AI Studio keys only for public toy repos.
# After edits: pytest -q. Do not git push. Refuse exploit/malware tasks.
# Prefer src/ + tests/. Never attach .env or credentials.json.

# auth concept (do not commit keys)
# export GEMINI_API_KEY=...          # AI Studio / Gemini API (student lab)
# # OR gcloud auth application-default login  # Vertex (enterprise lab)

# illustrative loop — confirm current command
# gemini "Explain how auth middleware works; cite file paths."
# gemini "Add GET /health + pytest matching existing FastAPI style (Vol. 18)."

CONTEXT_POLICY = {
    "max_files_explicit": 12,          # even with a huge window
    "forbid": [".env", "credentials.json", "*.pem"],
    "auth_sku_for_private_repo": "vertex",  # not consumer Gemini
    "human_approves_shell": True,
}

def attach_ok(path: str) -> bool:
    if any(path.endswith(s.replace("*", "")) or s.strip("*") in path
           for s in CONTEXT_POLICY["forbid"]):
        return False
    return True

When to Pick Gemini CLI

Pick Gemini CLI when

Vertex / Google Cloud is the mandated AI substrate.
Long-context repo or RFC tasks win your bake-off.
You want an open CLI you can inspect and script.
Multimodal “here is a screenshot of the bug” helps.

Pick something else when

Anthropic-standard shop → Claude Code.
JetBrains Tab + GitHub PRs → Copilot.
AI-native desktop UX → Cursor / Windsurf.
Zero-install web MVP → Bolt / Lovable / Replit.

Related Lectures

Lecture	Role
Google / Gemini catalog (22.1)	Models, Studio, Vertex vs this CLI
Gemini SDK	Service-side client, not the developer agent
Claude Code	Terminal bake-off peer
Context length	Why huge windows still need ignore lists
Agent loop / deployment	HITL + CI

Common Misconception

“Gemini CLI is the Gemini SDK.” SDK = your product; CLI = your laptop agent. Second: a 1M-token window means dump the monorepo including secrets. Third: AI Studio keys are fine for regulated customer repos. Fourth: Gemini CLI replaces Android Studio / VS Code Tab. Fifth: inventing free unlimited Vertex. Sixth: unsupervised gcloud prod deploys as a demo flex.

Knowledge Check

Short Answer: What is Gemini CLI’s primary shape? Answer: A terminal coding agent on Gemini/Vertex, not IDE ghost text.
True/False: Gemini CLI and the Vol. 18 Gemini SDK are interchangeable UXes. Answer: False.
Multiple Choice: Long context is best used to: (a) attach large specs/files deliberately, (b) paste .env, (c) skip tests. Answer: (a).
Short Answer: When should a private repo use Vertex instead of AI Studio keys? Answer: When enterprise IAM/DPA/VPC controls are required (prod/customer data).
True/False: Ignore lists still matter even with a huge context window. Answer: True.
Multiple Choice: Pricing posture is: (a) Gemini/Vertex usage (+ quotas), (b) a memorized $12 seat, (c) n8n Docker only. Answer: (a).
Short Answer: Name Claude Code’s role relative to Gemini CLI. Answer: Anthropic terminal-agent bake-off peer.
True/False: This lecture authorizes exploit generation via Gemini CLI. Answer: False.
Multiple Choice: Next lecture is: (a) Bolt.new, (b) Zapier, (c) Midjourney. Answer: (a).
Short Answer: Which Vol. 22.1 lecture catalogs Gemini/Vertex? Answer: Google / DeepMind / Gemini / Gemma / NotebookLM / Veo / Imagen / AI Studio / Vertex AI.

Key Takeaways

Gemini CLI is Google’s agentic terminal tool—pair with Vertex for private repos.
Long context is a strength, not a license to dump secrets or skip retrieval hygiene.
Do not confuse it with the Gemini SDK inside Vol. 18 services.
Bake off against Claude Code; keep Copilot/Cursor for Tab.
Next: Bolt.new — browser prompt-to-app.

Trainer’s Guide

Lab: Same /health + pytest task as Claude Code, on Gemini CLI (or a scripted mock if keys are unavailable). Compare cited files vs actual tree. Discuss Studio vs Vertex for a fake “PII repo.”

Transition: “What if the user is not a developer living in a terminal?” Preview Bolt.new WebContainers.

Recap: Gemini CLI closes the terminal-agent pair: Google long context + Vertex privacy SKUs, still HITL. Continue to Bolt.new.