💻 Copilot CLI

In one sentence

Not Copilot used without leaving your IDE, but Copilot that operates within the shell’s working context.

Key Features

A full-stack AI environment packed into a single terminal.

• Multi-file context awareness: Treats the entire repository as a single workspace.
• Code generation & editing: Make edits from the terminal, review diffs, and approve.
• Command execution: Run builds, tests, lint, and more — read the results and proceed.
• IDE-agnostic: Works anywhere there’s a terminal: VS Code, Vim, over SSH, etc.
• Skills / MCP support: Add the capabilities and external tools you need.
• Non-interactive execution: Run via CI, cron, or scripts with copilot -p "...".

CLI vs VS Code Chat

Powerful Built-in Agents

Copilot CLI includes standard agents for common tasks.

Rubber Duck — Cross-model Review

🦆 Experimental: A different model family from the main model provides a “second opinion” as a constructive critic, independently reviewing each phase of planning, implementation, and testing.

flowchart LR
  User([👤 Developer])
  Main["🧠 Main Model<br/>e.g. Claude"]
  Out["📝 Plan / Implementation / Test"]
  Duck["🦆 Rubber Duck<br/>e.g. GPT-5.4"]
  Final([✅ Verified output])
  User --> Main
  Main --> Out
  Out -->|"Independent review"| Duck
  Duck -->|"Blind spots / wrong assumptions"| Main
  Main --> Final

  classDef user fill:#0a1a14,stroke:#9bbc0f,color:#9bbc0f,stroke-width:2px
  classDef main fill:#0a0e27,stroke:#00f0ff,color:#00f0ff,stroke-width:2px
  classDef duck fill:#1a1500,stroke:#ffb000,color:#ffb000,stroke-width:2px
  classDef out fill:#1a0a2e,stroke:#ff2e88,color:#ff2e88,stroke-width:2px
  class User,Final user
  class Main main
  class Duck duck
  class Out out

Why does it work? ── When the same model checks its own output, it gets caught by the same assumptions and the same blind spots. A model from a different family has different training data and different values, so it can catch logic errors that were invisible to the first model.

Other Useful Commands

The CLI lets you instantly check models, sharing, experimental features, and environment info via slash commands.

Non-interactive Mode (Programmatic Execution) ★

🎯 If you find yourself asking the same thing in the interactive UI every day, that’s a signal to script it. Put it on cron with -p and let it run while you sleep.

Basic Usage

# Pass the prompt directly
copilot -p "Explain this file: ./complex.ts"

# Pipe it in
echo "Explain this file: ./complex.ts" | copilot

🔒 Minimize permissions. When running in CI, always use --allow-tool with a whitelist. Never use --allow-all. 💰 1 call = 1 Copilot premium request consumed. When making hundreds of batch calls, keep an eye on consumption via Copilot Metrics ↗.

Practical Automation Examples

Common Shell Script Patterns

Capture into a variable

node_version=$(copilot -p 'What Node.js version does this project require? Number only.' -s)
echo "Required: $node_version"

Use in a conditional

if copilot -p 'Does this project have TypeScript errors? Reply only YES or NO.' -s | grep -qi "no"; then
  echo "✅ Clean"
else
  echo "❌ Type errors detected" && exit 1
fi

Process multiple files sequentially

for file in src/api/*.ts; do
  echo "--- Reviewing $file ---" | tee -a review.md
  copilot -p "Review $file for error handling issues" -s \
    --allow-tool='shell(git:*)' | tee -a review.md
done

Using with GitHub Actions

- name: Generate test coverage report
  env:
    COPILOT_GITHUB_TOKEN: ${{ secrets.COPILOT_PAT }}
  run: |
    copilot -p "Run the test suite and produce a coverage summary" \
      -s --allow-tool='shell(npm:*), write' --no-ask-user

• 🔑 Pass authentication via environment variable. Priority order: COPILOT_GITHUB_TOKEN → GH_TOKEN → GITHUB_TOKEN
• 🧾 Use fine-grained PAT (v2) with the “Copilot Requests” permission (old ghp_ format is not supported)
• 📦 Can be called from cron / Jenkins / GitLab CI the same way (just pass auth via environment variables)

Design Tips

• 🎯 Be specific with prompts — Spell out file names, function names, and expected output format. Specifying the output format (e.g., “Number only.” or “Reply YES or NO.”) makes downstream parsing easier.
• 🛡️ Allow only necessary tools — Narrow the scope with --allow-tool='shell(git:*), write'
• 🔁 Think about idempotency — When running in Actions, limit to operations that are safe to run multiple times with the same input (have a human review commits/pushes)
• 📊 Keep logs — Save the entire session with --share='./session.md' so you can trace the reasoning behind results later
• 🧩 Build in interactive mode, then port to -p — The fastest path is to iterate in interactive mode until the prompt is solid, then script it with -p

📘 Details:

• Running GitHub Copilot CLI programmatically ↗
• Programmatic reference (full flag list) ↗
• Automating tasks with Copilot CLI and GitHub Actions ↗