GitHub Copilot CLI

What it is

The terminal descendant of GitHub Copilot. It installs as a global npm package, @github/copilot, and gives you an agent in the shell rather than in the editor — the same lineage as the autocomplete everyone knows, grown into something that reads a repository and makes changes.

Proprietary, from GitHub and Microsoft. It is bundled with the paid Copilot tiers rather than sold separately: Pro at roughly $10 a month, Pro+ at roughly $39, Max at roughly $100.

How the loop works

A conventional terminal agent loop, with two things that are not conventional.

First, model routing is automatic and cross-vendor. Requests are routed across the Claude, GPT and Gemini families depending on the work, which is a position no single-lab harness can take. You do not pick the model; GitHub does. Whether that is a feature depends entirely on whether you trust the routing — it removes a decision, and it removes a control.

Second, and more distinctive, is the GitHub context. This is the deepest available integration with pull requests, issues and review history, because the company that owns the agent owns the platform. Asking it about the discussion on a PR, or the history behind a contested file, is a native operation rather than an API integration you configured. For work that lives in that context, nothing else on this board is as well placed.

Isolation and permissions

Permissioned tools, in the now-standard pattern: the agent proposes, you approve. Session context includes your GitHub identity and the repository you are working in, which is convenient and also worth thinking about — the agent operates with your access to your organisation's code, and that reach is wider than a single working tree.

Who it is for

Teams that already pay for Copilot seats and already do everything in GitHub. The marginal cost of adopting it is close to zero and the integration is genuinely better than anything you could assemble yourself. If your review culture lives in PR threads, this is the agent that can read them.

Limits

Billing changed in May 2026 to token-based AI credits — 1,500 credits on Pro. That is a meaningful shift from a flat seat fee: heavy agent use now draws down a meter, and the cost of a long run is no longer bounded by your subscription price. Model your usage before assuming a Pro seat covers an agent workflow.

It is also not model-agnostic in the way that matters commercially. You can be routed to Claude, GPT or Gemini, but you pay GitHub, not the provider — so you get cross-vendor models without cross-vendor pricing leverage, and you cannot take your usage elsewhere. Being proprietary, there is no option to read the loop or run it yourself.

Alternatives on this board

• Claude Code — single-vendor, but you choose the model and pay the lab.
• Editor agents — including Copilot's in-editor form, for a tighter loop.
• Amp — the other proprietary option built around deep repository search.

Sources

• GitHub Copilot documentation
• Copilot plans and pricing
• @github/copilot on npm