Table of Contents
🌏 中文版
If your team already pays for GitHub Copilot, you already have a terminal agent — you may just not have opened it.
Copilot CLI entered public preview in September 2025 and went generally available on February 25, 2026, included in every Copilot plan, free tier included. This post covers whether it's worth switching to, and where its real differentiation lies.
The differentiator isn't the agent, it's the platform
Up front: plan mode, autopilot, subagents, hooks, skills, MCP — every other tool in this series has these. You can't tell them apart from a feature list. What sets Copilot CLI apart is that it lives on GitHub.
Concretely, three things:
A built-in GitHub MCP server. Not "it supports MCP so you can wire it up" — it ships configured. You can say "find the open issues related to this change" and it searches issues, reads labels and activity, and summarizes scope without you opening a browser. The objects it works on are issues, branches, and PRs, not just local files.
Org policies are inherited automatically. The CLI picks up your organization's existing Copilot governance — branch protections, required checks, and model availability allowlists all still apply. For a company already on Copilot Business or Enterprise, adopting a terminal agent doesn't mean a fresh security review. (The cost: on Business/Enterprise an admin has to enable Copilot CLI on the Policies page first; individuals can't turn it on themselves.)
Local and cloud are interchangeable. Prefix any prompt with & and the work is handed to the cloud-based Copilot coding agent in the background, freeing your terminal; /resume switches between local and remote sessions. You can also start a cloud agent session on github.com and pull it down locally to continue.
Installation and basics
npm install -g @github/copilot
copilotHomebrew, WinGet, a shell script, and standalone executables are also available; the Homebrew, WinGet, and install-script paths self-update. It runs on macOS, Linux, and Windows (PowerShell 6+), and ships in the default GitHub Codespaces image.
Run /init first to generate project-tailored Copilot instructions. Authentication uses your existing GitHub account, with OAuth device flow, GitHub CLI token reuse, and CI/CD-friendly GITHUB_ASKPASS also supported.
Three levels of autonomy
Shift+Tab cycles between modes:
| Mode | Behavior |
|---|---|
| Default | Every file-modifying or command-executing tool call needs approval; you can approve a tool for the whole session |
| Plan | Analyzes the request, asks clarifying questions, and produces a structured implementation plan before writing code |
| Autopilot | Runs to completion without step-by-step approval |
For full autonomy there's --allow-all / --yolo, but the more interesting control is /sandbox enable: it doesn't sandbox the CLI itself — it restricts what the commands and tools Copilot runs on your behalf can reach in your filesystem, network, and system capabilities. That distinction matters: the sandbox constrains the hands, not the head.
Built-in custom agents
Copilot CLI ships with a set of specialized subagents the model delegates to when it judges that worthwhile:
| Agent | Job |
|---|---|
| Explore | Fast codebase analysis without consuming your main context |
| Task | Runs tests and builds; brief summary on success, full output on failure |
| General purpose | Complex multi-step work in a separate context |
| Code review | Surfaces only genuine issues, minimizing noise |
| Research | Deep research across your codebase, related repos, and the web, with citations |
| Rubber duck | A constructive critic, consulted automatically by the CLI |
Rubber duck deserves its own note: it doesn't appear in the /agent picker — Copilot consults it on your behalf when it wants a second opinion. Shipping a built-in contrarian is the most unusual entry on that list.
/fleet goes the other direction: run the same task across multiple subagents in parallel, optionally on different models, then converge on one decision-ready result you choose to apply.
Custom agents are Markdown files (.agent.md) definable at the user level (~/.copilot/agents), repo level (.github/agents), or org/enterprise level (an /agents directory in the .github-private repo).
Billing: AI Credits
This is the easiest part to get wrong. Copilot bills in GitHub AI Credits, where 1 credit = $0.01 USD:
| Plan | Monthly | Included credits |
|---|---|---|
| Free | $0 | Limited; includes Copilot CLI and agent mode |
| Pro | $10/user | $15 |
| Pro+ | $39/user | $70 |
| Max | $100/user | $200 |
| Business / Enterprise | Separate | Admins set limits and decide whether overage is allowed |
Key points:
- Code completions and next edit suggestions don't consume credits and stay unlimited on paid plans. Chat, agent mode, code review, the cloud agent, and Copilot CLI do.
- Burn rate varies by model.
/usageshows credits spent this session, session duration, lines edited, and per-model token breakdown. - On Business/Enterprise, an admin decides whether hitting the cap pauses Copilot or bills onward.
Models span Anthropic, OpenAI, and Google; /model switches mid-session, reasoning effort is configurable, and Ctrl+T toggles whether reasoning is shown.
Context management
/context gives a visual overview of token usage and /compact compresses history manually. At roughly 95% of the context window it auto-compacts in the background without interrupting you.
There are also two memory layers: repository memory retains conventions and patterns for a codebase, and cross-session memory lets you ask about past work, files, and PRs.
Who it fits
Good fit:
- Developers whose org already pays for Copilot — marginal cost is zero and no new security review is required
- People whose work centers on issues and PRs, not just local files
- People who want local/cloud handoff —
&to send it out,/resumeto take it back - People whose editor has no official Copilot extension: the CLI is editor-independent
Poor fit:
- Teams outside the GitHub ecosystem — platform integration is the entire moat; remove it and this is an ordinary terminal agent
- People who want full control over models and providers — GitHub sets the model list and org admins can narrow it further
- Individuals inside a Business/Enterprise org who want to self-enable — you're waiting on an admin
Overall
The judgment here is simple: how much of your time is spent on GitHub? If your workflow is issue → branch → PR → review, this pulls that line into your terminal using money you've already spent. If GitHub isn't central to your work, it offers no clear advantage over Claude Code or OpenCode.
It forms an interesting contrast with Codex: both are terminal agents riding an existing subscription, but Codex rides ChatGPT, a consumer subscription, while Copilot CLI rides a developer platform the company already bought. The latter is usually an easier sell internally.
References
Loading...