What Gemini CLI is documented to be good at

Five things Gemini CLI is documented to be good at#

These come from the official README, the geminicli.com docs, and worked examples in the repo. None are taste-based — each maps to a feature the CLI specifically ships.

1. Codebase analysis using the 1M-token context window#

Gemini’s 1M-token window means you can ask broad cross-cutting questions and let the agent pull in whatever it needs without careful file pre-selection:

• “What does this monorepo do, end to end?”
• “Find every place we touch the Stripe API and tell me if we handle the customer-deleted webhook correctly.”
• “This Rails app is upgrading to Rails 7.2 — what’s likely to break?”

The model can read large slices of the repo in one pass, which means fewer “I’ll need to see the related files” round-trips.

2. Wiring Gemini CLI to your tools via MCP#

Gemini’s MCP client implementation is one of its strongest features. Add a few lines to ~/.gemini/settings.json and the agent can:

• Browse your GitHub repos and issues (via the GitHub MCP server)
• Query a database (Postgres, MongoDB, BigQuery, Firestore — via the MCP Toolbox)
• Drive a browser (Playwright MCP)
• Read Figma designs (Figma MCP)
• Touch your filesystem with explicit allowlists (filesystem MCP)

{
  "mcpServers": {
    "github": {
      "command": "path/to/github-mcp-server",
      "env": { "GITHUB_TOKEN": "$GITHUB_TOKEN" },
      "trust": false
    }
  }
}

A real shape: “Open issue #142 in this repo, read the linked design in Figma, then implement the changes.” — three MCP servers cooperating.

See GCL.6 MCP integration for the full mechanics.

3. Headless scripting and CI runs#

The -p flag (or any non-TTY invocation) flips Gemini CLI into non-interactive mode. Output formats:

• Plain text (default)
• --output-format json — single JSON object
• --output-format stream-json — JSONL events, one per line, suitable for piping

# CI-friendly run with structured output
gemini -p "Summarise the changes in the last commit" \
       --output-format json > summary.json

Common CI shapes from the docs:

• Generate release notes from git log
• Sanity-check CHANGELOG.md against actual commits
• Triage open issues against a labelling rubric

4. The official GitHub Action — PR review and on-demand @gemini-cli#

The google-github-actions/run-gemini-cli Action lets you wire Gemini CLI into a workflow:

• Pull-request review on every PR — comments with line-level suggestions
• Issue triage — auto-label and assign on open
• On-demand mention — comment @gemini-cli "fix the failing test" on an issue or PR

- uses: google-github-actions/run-gemini-cli@v1
  with:
    prompt: "Review this PR for security and correctness issues."
    gemini-api-key: ${{ secrets.GEMINI_API_KEY }}

5. Prototyping against the free tier#

This is the use case Gemini CLI’s economics tilt towards. With 1,000 requests / day on a personal Google account at no cost, you can:

• Try a refactoring approach without burning tokens
• Spike a side project end-to-end without paying anything
• Explore a new framework’s docs interactively
• Use it as a daily learning companion

Compare with Claude Code (no free tier) and Codex CLI (Free tier is explore-only; productive use needs a paid plan (Go/Plus/Pro) or an OpenAI API key — the old $5 trial credit is gone): for tinkerers and learners, Gemini CLI is the lowest-friction starting point.

What it’s not (yet) good at#

Honest counterweight — areas where the docs and issue tracker are clear about gaps:

• Multi-agent / subagents. No native primitive. Issue #3132 tracks this. Mono-agent design means long sessions accumulate noise; Codex CLI and Claude Code both have subagent support.
• Local model support. No Ollama, LM Studio, vLLM. Issue #5938 is the placeholder. Hard blocker for HIPAA / SOX / GDPR / air-gapped environments.
• Sycophancy. Issue #4556 — model is excessively agreeable, fills context with “you are 100% correct,” ignores “be brutally honest” instructions. Real complaint from real users; no setting to tune it down yet.
• Tool loops. Issue #1531 — when a tool fails (write, shell, search), the model often retries indefinitely instead of surfacing the error cleanly.

If you need any of those, look at Codex CLI (sub-agents, sandboxed-by-default) or Claude Code (subagent spawning).

Realistic first session — the build is failing#

A close-to-canonical first session, suitable for a Tuesday afternoon:

cd ~/projects/my-broken-repo
gemini

Then type:

the build is failing — figure out why and fix it

The agent reads the error, opens config files, infers the cause (deprecated flag, version mismatch, missing dependency), edits the file, re-runs the build, and reports back. Default behaviour asks before running shell commands; reading and editing files inside the workspace happens without prompting.

If you want extra safety, turn the sandbox on:

gemini --sandbox

(See GCL.5 Pitfalls on why the sandbox is OFF by default and what changes when you turn it on.)

What’s next#

• §GCL.5 Pitfalls — the rough edges and known issues
• §GCL.6 MCP integration — adding MCP servers
• §GCL.3 Auth — picking the right auth path for your use case