GitHub Copilot CLI & Coding Agent: The Terminal-Native AI Developer

GitHub Copilot has evolved from autocomplete into a three-pronged AI development platform — and the coding agent alone has already authored 1 million pull requests in its first 5 months. September 2025 brought the biggest expansion yet: Copilot CLI in public preview, Copilot Coding Agent going general availability, Copilot Spaces hitting GA, and GPT-5-Codex rolling out to Copilot users.

This is what’s actually available now, how it compares to Cursor and Claude Code, and where Copilot fits in the AI coding landscape.

---

The Three Surfaces

Copilot CLI (Public Preview, Sept 2025)

Terminal-native AI agent. No IDE required — work with Copilot directly in your command line.

How it works:

npm install -g @github/copilot
copilot auth login  # authenticate with GitHub account

Key capabilities:

• Build, edit, debug, and refactor code with a terminal AI agent
• GitHub integration out of the box — repos, issues, PRs via natural language
• Built-in GitHub MCP server (repos, issues, PRs, CI/CD, code quality, security)
• Custom MCP server support for extensibility
• Three approval modes: read-only (explicit approvals), auto (full workspace access), full access (read anywhere, run commands with network)
• Preview every action before execution — nothing happens without confirmation
• Attach images (screenshots, wireframes) for design context

Use cases:

• Implement a feature from a GitHub issue
• Debug a failing test locally
• Refactor code across multiple files
• Explore a new codebase
• Review and address PR comments

---

Copilot IDE Extension

The VS Code / VS Code fork integration with three modes:

Mode	What It Does
Ask	Q&A about your codebase, get explanations
Edit	Multi-file editing with AI assistance
Agent	Autonomous task execution with tool access

Model picker (Sept 2025):

• Claude Sonnet 4.5
• GPT-5-Codex (agentic coding-optimized)
• Gemini models
• Claude Fable 5.1

---

Copilot Coding Agent (GA, Sept 2025)

The asynchronous, autonomous agent. Delegate a task, and Copilot opens a draft PR and works in its own environment via GitHub Actions.

How it works:

1. Assign Copilot to an issue (or use “Delegate to coding agent” in VS Code)
2. Copilot plans the implementation in its own development environment
3. Opens a draft pull request
4. You review, leave comments, and Copilot iterates

What it handles:

• Bug fixes with clear reproduction steps
• Dependency updates and maintenance
• Test coverage expansion
• Documentation improvements
• Refactoring with well-defined scope

By the numbers (GitHub Octoverse 2025):

• 1 million+ PRs authored in 5 months (May–Sept 2025)
• 72.6% of code review users report improved effectiveness
• PR merge rate: +23% YoY
• Commits: +25% YoY (986M total, 100M in August alone)

---

GitHub MCP Server (Built In)

Copilot ships with the GitHub MCP Server — giving AI tools direct access to GitHub’s platform.

Available toolsets:

Toolset	Capabilities
repos	Browse, query code, search files, analyze commits
issues	Create, update, manage issues
pull_requests	Create, review, manage PRs
actions	GitHub Actions workflows and CI/CD
code_security	Security findings, Dependabot alerts
code_quality	Code scanning, quality analysis
dependabot	Dependency vulnerability management
secret_protection	Secret scanning
copilot	Copilot-specific tools (including coding agent)
copilot_spaces	Spaces for grounded context

32k+ stars on GitHub. Extensible via custom MCP servers in .github/copilot.md.

---

Copilot Spaces (GA, Sept 2025)

Ground Copilot in curated project context so responses follow your team’s standards.

What you add to a space:

• Files (source code, docs, configs)
• Pull requests
• Issues
• Repositories
• Custom instructions

Result: Copilot understands your systems, standards, and open tasks — not just generates generic code.

---

GPT-5-Codex in Copilot

OpenAI’s GPT-5-Codex (agentic coding model) is rolling out to Copilot Pro, Pro+, Business, and Enterprise users via the model picker in VS Code.

Key features:

• Trained for real-world software engineering
• Code review capability: matches PR intent to actual diff, reasons over codebase, executes tests
• Available in ask, edit, and agent modes
• Sandboxed execution (network disabled by default)
• Three approval modes: read-only, auto, full access

Rollout: Gradual. Requires VS Code v1.104.1+. Enterprise admins must enable the GPT-5-Codex policy.

---

How It Compares

Dimension	GitHub Copilot CLI	Cursor	Claude Code
Surface	Terminal + IDE	IDE (VS Code fork)	Terminal
GitHub Integration	Native (built-in MCP)	None native	Via extensions
Async Agent	Yes (GitHub Actions)	No	No (sync only)
Model Choice	Claude, GPT, Gemini	Claude, GPT, Gemini	Claude only
Pricing	$10-$39/mo	$20/mo	Free (API billing)
IDE Lock-in	VS Code / forks	Custom IDE	Terminal (any IDE)
Context Window	Copilot Spaces + files	Full codebase	Conversation-based
Code Review	Built-in (GPT-5-Codex)	Basic	Via external tools

Bottom line: Copilot CLI is best if you live in the GitHub ecosystem. Cursor is best for IDE-first developers who want full codebase awareness. Claude Code is best for terminal-first developers who want Claude’s reasoning.

---

Pricing

Plan	Price	Copilot CLI	Coding Agent	Spaces
Copilot Free	$0	❌	❌	Limited
Copilot Pro	$10/mo	✅	✅	✅
Copilot Pro+	$39/mo	✅	✅ (priority)	✅
Copilot Business	$19/user/mo	✅	✅	✅
Copilot Enterprise	$39/user/mo	✅	✅	✅

All paid plans include the coding agent. Usage scales with plan tier (Plus and Enterprise get more agent minutes).

---

Key Takeaways

1. Three surfaces for different workflows — CLI for terminal, IDE for editing, coding agent for async delegation
2. 1M+ PRs in 5 months — the coding agent is the fastest-growing AI coding tool by adoption
3. GitHub MCP Server is the killer feature — built-in repos, issues, PRs, CI/CD access no other tool matches
4. Copilot Spaces solve the context problem — ground AI in your project’s actual files and conventions
5. GPT-5-Codex integration — OpenAI’s best coding model, now available alongside Claude in Copilot
6. MCP extensibility — custom MCP servers make Copilot endlessly extensible
7. Approval modes — read-only, auto, full access let you choose your risk level
8. Best for GitHub-native workflows — if you live in GitHub, nothing else comes close

---

Resources

• GitHub Copilot
• Copilot CLI Documentation
• GitHub MCP Server
• Copilot Spaces
• Copilot Pricing
• GitHub Octoverse 2025 Report
• Copilot CLI Changelog
• Copilot Coding Agent Changelog
• GPT-5-Codex in Copilot

---

Research date: 2025-09-18. Data verified against GitHub changelog, GitHub Octoverse 2025 report, and official Copilot documentation. Verify latest pricing and features at github.com/pricing.