跳到主要内容Skip to main content
← Coding

GitHub Copilot CLIGitHub Copilot CLI

GitHub
编程Coding 终端Terminal 本地Local 多模型Multiple models 全流程End-to-end partial lifecycle: active 核验Checked 2026-10-23

Copilot 在终端里的编码 agent 形态,与 GitHub 订阅同源。

它能在终端里走完一条链。

"读仓库 / 改代码 / 跑测试 / 开 PR / 管 Issue" 都在终端里完成。

官方 README 说它

「Powered by the same agentic harness as GitHub's Copilot coding agent」

它与既有 GitHub Copilot(IDE 形态)是同一产品的两种装法。

本条目只写 CLI 的差异点:

  • 程序化入口
  • 终端级审批
  • 按提示词扣 premium requests 的额度扣减方式

GitHub Copilot CLI is the Copilot coding agent that runs in your terminal, carried by the same GitHub subscription as the IDE form factor. The official README says it is "Powered by the same agentic harness as GitHub's Copilot coding agent." Three things differ from the IDE form. It has a programmatic entry point (copilot -p / --prompt) that can go into scripts and CI. Approval is terminal-level: three choices on first use of a writing or executing tool, plus --allow-all-tools / --allow-tool / --deny-tool for unattended runs, and some official docs say denial wins. Billing is per prompt too: "Each time you submit a prompt to Copilot CLI, your monthly quota of Copilot premium requests is reduced by one."

适合与不适合When to use it, when not to

适合三类人:

  • 已在 GitHub 生态、想在终端把「读仓库 → 改代码 → 跑测试 → 开 PR / 管 Issue」一条链走完的人
  • 需要脚本化 / headless 的人(copilot -p 配审批选项,可进 CI)
  • 需要 MCP / 插件 / 自定义 agent 生态的人

不适合不想付 Copilot 订阅的人:CLI 随订阅、不能单独买。

也不适合想「全自动又零风险」的人:自动批准必须与隔离环境配套。官方对此有明确风险提示。

People already in the GitHub ecosystem who want to walk the whole chain from reading a repo to opening a pull request in the terminal. The official use case gives tasks such as "create a pull request on GitHub.com".

People who need scripted or headless use: copilot -p with approval flags is CI-friendly. Anyone who wants the MCP / plugin / custom-agent surface.

Not a fit for people who don't want a Copilot subscription, because the CLI ships with the subscription and cannot be bought on its own. Nor for anyone who wants "fully autonomous and zero risk": auto-approval has to be paired with an isolated environment, and the official docs warn about that plainly.

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

出品属地Origin
国外出品foreign-made 国外主体(网信办公示名单未见对应备案主体)。foreign entity (no matching filing entity in the CAC registry).
国内使用Direct access
大陆可直连direct OK docs.github.com 直连实测 200。docs.github.com measured 200 direct.
备案状态CAC filing
未见于备案名单not in filing list:未见于网信办公示名单(2026-10-20 检索 13 批附件,https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm);该产品未面向大陆公众提供服务,备案不适用。not in the CAC public registry (13 batch attachments searched 2026-10-20, https://www.cac.gov.cn/2024-04/02/c_1713729983803145.htm); the product does not serve the mainland public, so filing does not apply.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准(按批次滚动更新,本站只抄编号不解读);「大陆可直连」是本站从大陆网络直连实测的结果,不是厂商承诺。GitHub 类站点当日实测有波动,口径写在依据里。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise. GitHub-hosted sites fluctuated on the measurement day — the note says so.

固定坐标系Fixed coordinate system

8 个维度dimensions, 与同分区其他对象逐项可比each comparable item by item against others in the same section.

模型与开放条件Model & openness

多厂商模型,可在 CLI 内切换。

官方 README 原文:

官方原文:"By default, copilot utilizes Claude Sonnet 4.5. Run the /model slash command to choose from other available models, including Claude Sonnet 4 and GPT-5"

官方 changelog 写 GA 时的表述是:

官方原文:"Choose from the latest models from Anthropic, OpenAI, and Google"

注意官方两页对默认模型的表述不同:

  • README 写 Claude Sonnet 4.5
  • about 页写「The default model used by GitHub Copilot CLI is Claude Sonnet 4. GitHub reserves the right to change this model」

具体可用模型清单随套餐变化,完整清单见 Copilot 模型页(与 IDE 条目 同一套账号体系)。

第三方 agent 委派沿用 Copilot 平台能力(Preview)。

Multi-vendor models, switchable inside the CLI.

Official README wording:

Official wording: "By default, copilot utilizes Claude Sonnet 4.5. Run the /model slash command to choose from other available models, including Claude Sonnet 4 and GPT-5"

The official changelog's GA wording was:

Official wording: "Choose from the latest models from Anthropic, OpenAI, and Google"

Note the two official pages describe the default model differently:

  • the README says Claude Sonnet 4.5
  • the about page says "The default model used by GitHub Copilot CLI is Claude Sonnet 4. GitHub reserves the right to change this model"

The exact list of available models varies by plan; the full list is on the Copilot models page (same account system as the IDE entry).

Third-party agent delegation follows Copilot platform capabilities (Preview).

运行位置Where it runs

本地终端进程,生命周期绑定终端会话。

官方支持的系统:

  • Linux
  • macOS
  • Windows:走 Powershell 或 Windows Subsystem for Linux (WSL)

官方原文写支持:

官方原文:"Linux / macOS / Windows from within Powershell and Windows Subsystem for Linux (WSL)"

about 页另注:

官方原文:"Native Windows support in Powershell is available, but experimental"

另有一条云端运行面。 copilot --cloud 可在隔离的云托管环境里起会话 (官方文档写 Cloud sandboxes「in public preview」,其策略「inherit from Copilot cloud agent policies」,可跨机器续会话、并行跑任务)。

这与本地终端会话是两个运行面,不要混为一谈。

A local terminal process, lifecycle bound to the terminal session.

Officially supported systems:

  • Linux
  • macOS
  • Windows: via Powershell or Windows Subsystem for Linux (WSL)

Official wording on support:

Official wording: "Linux / macOS / Windows from within Powershell and Windows Subsystem for Linux (WSL)"

The about page adds:

Official wording: "Native Windows support in Powershell is available, but experimental"

There is also a cloud execution surface. copilot --cloud starts a session in an isolated cloud-hosted environment (official docs say cloud sandboxes are "in public preview", their policies "inherit from Copilot cloud agent policies", sessions can continue across machines and run tasks in parallel).

These are two execution surfaces — the local terminal session and the cloud one; do not conflate them.

本地文件Local files

以「信任目录」为界。 启动会话时,官方要求你确认信任「confirm that you trust the files in, and below, the directory from which you launched the CLI」。

这个受托目录控制 "where Copilot CLI can read, modify, and execute files"。

关键限定:官方明写

官方原文:"Scoping of permissions is heuristic and GitHub does not guarantee that all files outside trusted directories will be protected"

别把信任目录当成强沙箱。

想要更硬的边界,可开本地沙箱:官方原文:"run /sandbox enable inside a session", 用以限制文件系统、网络与系统能力。或者用 copilot --cloud 把整个会话放进云隔离环境。

Bounded by "trusted directories". When starting a session, the official docs require you to "confirm that you trust the files in, and below, the directory from which you launched the CLI".

This trusted directory controls "where Copilot CLI can read, modify, and execute files".

Key caveat: the official docs state plainly

Official wording: "Scoping of permissions is heuristic and GitHub does not guarantee that all files outside trusted directories will be protected"

Do not treat trusted directories as a hard sandbox.

For a harder boundary, enable the local sandbox: Official wording: "run /sandbox enable inside a session", which restricts filesystem, network and system capabilities. Or use copilot --cloud to put the whole session in a cloud-isolated environment.

关机后的任务Tasks after shutdown

有明确的后台委派与云端会话。 官方 changelog 原文:

官方原文:"Prefix any prompt with & to delegate work to the Copilot coding agent in the cloud, freeing your terminal for other tasks. Use /resume to switch between local and remote coding agent sessions seamlessly"

也就是说,在终端里把任务甩给云端 coding agent,终端继续干别的。

云沙箱(public preview)另支持:

官方原文:"keep a session's state between uses, continue a session from a different machine, or run multiple tasks in parallel"

本地会话本身仍是随终端结束的。 它是客户端常驻型,真后台来自上面这条云端通道。

There is explicit background delegation and cloud sessions. Official changelog wording:

Official wording: "Prefix any prompt with & to delegate work to the Copilot coding agent in the cloud, freeing your terminal for other tasks. Use /resume to switch between local and remote coding agent sessions seamlessly"

That is, in the terminal you hand the task to the cloud coding agent and keep using the terminal for other things.

Cloud sandboxes (public preview) additionally support:

Official wording: "keep a session's state between uses, continue a session from a different machine, or run multiple tasks in parallel"

The local session itself still ends with the terminal. It is a client-resident type; the real background comes from the cloud channel above.

工具与扩展Tools & extensions

MCP 是默认带的能力。 官方 README 原文:

官方原文:"the coding agent ships with GitHub's MCP server by default and supports custom MCP servers"

添加方式:/mcp add 交互式添加;/mcp search 浏览 GitHub MCP Registry 直接安装(2026-06-23 changelog)。

官方文档另列的扩展面:

  • Custom instructions
  • Custom agents —— 含内置 Explore / Task / Code Review / Plan 等专职 agent,可并行
  • Hooks
  • Skills
  • Plugins

还有 LSP:需自行装 language server,配 ~/.copilot/lsp-config.json 或 .github/lsp.json。

支持 ACP 接入(官方概念页有「Use Copilot CLI via ACP」小节,changelog 提到 ACP 的 closeSession)。

local 会话在 macOS / Windows 还可开 computer use 操作桌面应用。

MCP is a default capability. Official README wording:

Official wording: "the coding agent ships with GitHub's MCP server by default and supports custom MCP servers"

How to add: /mcp add for interactive adding; /mcp search browses the GitHub MCP Registry for direct install (2026-06-23 changelog).

Other extension surfaces listed in official docs:

  • Custom instructions
  • Custom agents — including built-in Explore / Task / Code Review / Plan specialist agents, runnable in parallel
  • Hooks
  • Skills
  • Plugins

There is also LSP: you install language servers yourself, configured via ~/.copilot/lsp-config.json or .github/lsp.json.

ACP integration is supported (the official concepts page has a "Use Copilot CLI via ACP" section; the changelog mentions ACP closeSession).

On macOS / Windows, local sessions can additionally enable computer use to operate desktop apps.

上下文与记忆Context & memory

自动压缩 + 会话管理 + 仓库记忆三件套。

自动压缩:官方原文写会话接近 95% token 上限时「Copilot automatically compresses your history in the background」,以实现「virtually infinite sessions」。

手动操作:/compact 手动压缩;/context 查看 token 占用。

回滚用 /rewind。changelog 写它

官方原文:"no longer requires git and restores only the files Copilot changed, skipping any file whose contents no longer match what Copilot last wrote"

并可选 官方原文:"conversation-only 或 conversation + files"。

跨会话记忆沿用 Copilot Memory(与本分区 IDE 条目 同一功能,存仓库级约定与偏好)。

The trio: auto-compaction + session management + repository memory.

Auto-compaction: official docs say when a session nears the 95% token limit, "Copilot automatically compresses your history in the background", enabling "virtually infinite sessions".

Manual operations: /compact compacts manually; /context shows token usage.

Rollback uses /rewind. The changelog says it

Official wording: "no longer requires git and restores only the files Copilot changed, skipping any file whose contents no longer match what Copilot last wrote"

and optionally Official wording: "conversation-only or conversation + files".

Cross-session memory follows Copilot Memory (the same feature as the IDE entry in this section, storing repo-level conventions and preferences).

权限与限制Permissions & limits

审批是本条目最值得记的一维,颗粒度比 IDE 形态更细。

首次使用可写 / 可执行工具时,官方给三个选项:

  • 1. Yes
  • 2. Yes, and approve TOOL for the rest of the running session
  • 3. No, and tell Copilot what to do differently (Esc)

官方原文:

官方原文(审批提示的三个选项):"1. Yes" / "2. Yes, and approve TOOL for the rest of the running session" / "3. No, and tell Copilot what to do differently (Esc)"

选项 2 的含义是本会话内该类命令全部放行。 官方举例:批准 rm ./this-file.txt 后,本会话可跑任意 rm。

程序化 / 无人值守模式有三个开关:

  • --allow-all-tools —— 任何工具都不问
  • --allow-tool —— 放行指定工具
  • --deny-tool —— 禁止指定工具,优先级最高

取值支持「shell(COMMAND)、write、MCP_SERVER_NAME」,且 --deny-tool 优先于 --allow-tool / --allow-all-tools。

官方对自动批准的警告值得原样记住:

官方原文:"Copilot has the same access as you do to files on your computer, and can run any shell commands that you can run, without getting your prior approval"

官方并建议在 VM / 容器 / 无网环境里用自动批准,以限制影响面。

内容排除方面,Business / Enterprise 用户的 CLI:

官方原文:"respects content exclusion policies configured at the enterprise, organization, and repository levels"

企业与组织策略可整体禁用 CLI。

Approval is the axis most worth remembering in this entry, with finer granularity than the IDE form.

On first use of write / executable tools, the official docs give three options:

  • 1. Yes
  • 2. Yes, and approve TOOL for the rest of the running session
  • 3. No, and tell Copilot what to do differently (Esc)

Official wording:

Official wording (the three approval-prompt options): "1. Yes" / "2. Yes, and approve TOOL for the rest of the running session" / "3. No, and tell Copilot what to do differently (Esc)"

Option 2 means all commands of that tool class are allowed for this session. Official example: after approving rm ./this-file.txt, any rm can run in the session.

Programmatic / unattended mode has three switches:

  • --allow-all-tools — never ask for any tool
  • --allow-tool — allow specified tools
  • --deny-tool — forbid specified tools, highest priority

Values support "shell(COMMAND), write, MCP_SERVER_NAME", and --deny-tool takes precedence over --allow-tool / --allow-all-tools.

The official warning about auto-approval is worth remembering verbatim:

Official wording: "Copilot has the same access as you do to files on your computer, and can run any shell commands that you can run, without getting your prior approval"

The official docs also recommend using auto-approval inside VMs / containers / offline environments to limit blast radius.

For content exclusion, the CLI for Business / Enterprise users:

Official wording: "respects content exclusion policies configured at the enterprise, organization, and repository levels"

Enterprise and organization policies can disable the CLI entirely.

适合什么任务Best for

适合三类人:

  • 已在 GitHub 生态、想在终端把「读仓库 / 改代码 / 跑测试 / 开 PR / 管 Issue」走完的人。官方用例直接给了「create a pull request on GitHub.com」这类任务
  • 需要脚本化 / headless 的人(copilot -p 配合审批选项,可进 CI)
  • 需要 MCP / 插件 / 自定义 agent 生态的人

不适合两类人:

  • 不想付 Copilot 订阅的人 —— CLI 随订阅,不能单独买
  • 想要「全自动又零风险」的人 —— 自动批准要与隔离环境配套,官方明确提示风险

Fits three kinds of people:

  • those already in the GitHub ecosystem who want to finish "read the repo / edit code / run tests / open PRs / manage issues" from the terminal — the official use cases literally include tasks like "create a pull request on GitHub.com"
  • those who need scripting / headless (copilot -p with approval options can run in CI)
  • those who need the MCP / plugins / custom-agent ecosystem

Not a fit for two kinds of people:

  • those unwilling to pay for a Copilot subscription — the CLI comes with the subscription and cannot be bought separately
  • those wanting "fully automatic yet zero risk" — auto-approval must be paired with an isolated environment; the official docs flag the risk explicitly

头号误解Top misconception

  • 以为 Copilot CLI 只做补全问答——官方是完整 coding agent(计划/编辑/测试/开 PR)Thinking Copilot CLI is a separate purchase — it comes with a Copilot subscription and has no standalone price
  • 以为 -p 非交互默认能改文件——需显式加审批选项(如 --allow-tool)Thinking the CLI costs no quota — each prompt deducts one premium request, and premium models multiply it
  • 以为信任目录能强隔离——官方说该边界是启发式、不保证Taking the about page's plan list (Pro/Pro+/Business/Enterprise) as the only wording — the quickstart page says all plans

未知项清单Known unknowns

  • CLI 与各 Copilot 档位的准确对应(Free 是否含 CLI)——官方未说明(已查 about-copilot-cli 与 cli-quickstart 两页,两页口径不一致)
  • 云沙箱(copilot --cloud)的并发、时长与存储上限(public preview)
  • 交互式与 programmatic 两种模式下 premium request 扣减的细化差异
  • ACP 接入的完整能力面(官方概念页列有该小节,细则未展开)
  • The exact mapping between the CLI and each Copilot plan (does Free include the CLI) — not stated officially (both about-copilot-cli and cli-quickstart were checked; the two pages are inconsistent)
  • Concurrency, duration and storage limits of the cloud sandbox (copilot --cloud) (public preview)
  • Fine-grained differences in premium-request deduction between interactive and programmatic modes
  • The full capability surface of ACP integration (the official concepts page lists the section but does not elaborate)

证据来源Evidence sources

判断可回到以下一手源复核。本站核验日Every judgement can be rechecked against the primary sources below. Verified on 2026-10-23,内容更新日content updated on 2026-10-08。

类型Type名称Name链接Link
docsGitHub · 关于 GitHub Copilot CLI(模式、审批选项、信任目录、模型与 premium requests)GitHub · About GitHub Copilot CLI (modes, approval options, trusted directories, models and premium requests)https://docs.github.com/en/copilot/concepts/agents/about-copilot-cli
docsGitHub · Copilot CLI 概念页(沙箱、上下文管理、自定义、内容排除、ACP)GitHub · Copilot CLI concepts page (sandbox, context management, customization, content exclusion, ACP)https://docs.github.com/en/copilot/concepts/copilot-surfaces/copilot-cli
docsGitHub · GitHub Copilot CLI 快速开始(npm / WinGet / Homebrew 安装、-p 非交互)GitHub · GitHub Copilot CLI quickstart (npm / WinGet / Homebrew install, -p non-interactive)https://docs.github.com/en/copilot/get-started/cli-quickstart
docsGitHub · 安装 GitHub Copilot CLI(各平台命令与前置条件)GitHub · Install GitHub Copilot CLI (per-platform commands and prerequisites)https://docs.github.com/en/copilot/how-tos/copilot-cli/set-up-copilot-cli/install-copilot-cli
docsGitHub · Requests in GitHub Copilot(premium requests 与模型倍率)GitHub · Requests in GitHub Copilot (premium requests and model multipliers)https://docs.github.com/en/copilot/concepts/billing/copilot-requests
changelogGitHub · Copilot CLI 离线 changelog(1.0.93 @ 2026-10-07)GitHub · Copilot CLI offline changelog (1.0.93 @ 2026-10-07)https://github.com/github/copilot-cli/blob/main/changelog.md
repoGitHub · copilot-cli 官方仓库(README:默认模型、MCP、LSP、PAT 认证)GitHub · copilot-cli official repository (README: default model, MCP, LSP, PAT auth)https://github.com/github/copilot-cli
changelogGitHub · Copilot CLI 正式可用(2026-02-25 GA,plan / autopilot / background delegation)GitHub · GitHub Copilot CLI generally available (2026-02-25 GA, plan / autopilot / background delegation)https://github.blog/changelog/2026-02-25-github-copilot-cli-is-now-generally-available/
changelogGitHub · Changelog(Copilot 标签页)GitHub · Changelog (Copilot label page)https://github.blog/changelog/label/copilot/

实测记录Measured runs

本档未跑端到端编码任务;能力结论来自官方文档,编码向测量协议见 protocols/coding/_protocol.md。

实测建议:重点测三件事。

  • 三选项审批中「本会话放行」的实际影响面(官方举例说明它会放大到整类命令)
  • copilot -p 在 CI(无 TTY)下的退出码与输出
  • 前缀 & 的后台委派是否真能把任务交给云端并 /resume 拉回

This site has not completed measured runs. The test protocol is at protocols/coding/_protocol.md.

Suggested measurement: focus on three things.

  • the real blast radius of "approve for the rest of the session" in the three-option prompt (the official example shows it expands to the whole command class)
  • exit codes and output of copilot -p under CI (no TTY)
  • whether the & prefix delegation really hands the task to the cloud and pulls it back via /resume

与 IDE 形态的差别(只说差异)Differences from the IDE form (differences only)

维度IDE 形态(见 ./copilot.md)CLI 形态(本条目)
入口编辑器 / 网站终端 copilot
非交互一般无copilot -p / --prompt,可进脚本与 CI
审批编辑器模式终端三选项 + --allow/--deny-tool 开关
云端运行面网站 / coding agentcopilot --cloud 云沙箱(public preview)
后台委派派 issue / 评论触发前缀 & 把任务交给云端 coding agent

共用的部分不重复写:模型清单、Copilot Memory、组织策略,均以 IDE 条目 为准;档位报价一律看官方 plans 页。

DimensionIDE form (see ./copilot.md)CLI form (this entry)
Entryeditor / websiteterminal copilot
Non-interactivegenerally nonecopilot -p / --prompt, usable in scripts and CI
Approvaleditor modesterminal three-option prompt + --allow/--deny-tool switches
Cloud execution surfacewebsite / coding agentcopilot --cloud cloud sandbox (public preview)
Background delegationdispatch via issue / comment triggerprefix & hands the task to the cloud coding agent

Shared parts are not repeated: model lists, Copilot Memory, org policies all follow the IDE entry; plan pricing always goes to the official plans page.

程序化与安装Programmatic use and installation

安装以 npm 包为主。官方前置条件是 Node.js 22 or later:

npm install -g @github/copilot

另有 WinGet(winget install GitHub.Copilot)、Homebrew(brew install --cask copilot-cli)与安装脚本。

认证除 /login 外,支持细粒度 PAT + 「Copilot Requests」权限,放进 GH_TOKEN 或 GITHUB_TOKEN。

官方示例:copilot -p "Show me this week's commits and summarize them" --allow-tool='shell(git)'

加 -s 只输出回答、省去用量信息。

Installation is primarily via the npm package. The official prerequisite is Node.js 22 or later:

npm install -g @github/copilot

There are also WinGet (winget install GitHub.Copilot), Homebrew (brew install --cask copilot-cli) and an install script.

Besides /login, authentication supports a fine-grained PAT with the "Copilot Requests" permission, placed in GH_TOKEN or GITHUB_TOKEN.

Official example: copilot -p "Show me this week's commits and summarize them" --allow-tool='shell(git)'

Add -s to output only the answer, omitting usage info.

权限与审批(本条目最值得记的一维)Permissions and approvals (the axis most worth remembering)

三选项审批(官方原文):

1. Yes
2. Yes, and approve TOOL for the rest of the running session
3. No, and tell Copilot what to do differently (Esc)

选项 2 的含义比字面更大。官方举例:批准 rm ./this-file.txt 后,本会话可跑任意 rm(如 rm -rf ./*)。这是「按工具类别放行」,不是逐条命令放行。

无人值守用三个开关,且 deny 优先:

开关作用
--allow-all-tools任何工具都不问
--allow-tool放行指定工具
--deny-tool禁止指定工具,优先级最高

取值支持 shell(COMMAND)(可到 shell(git push) 这一级)、write、MCP_SERVER_NAME。

官方警告:用 --allow-all-tools 时,它与你有同等本机访问权。

本机文件与命令都算,而且不再逐条确认。

官方建议在 VM / 容器 / 无网环境里使用,以限制影响面。

信任目录不是强沙箱。官方明写其边界

「is heuristic and GitHub does not guarantee that all files outside trusted directories will be protected」

The three-option approval prompt (official wording):

1. Yes
2. Yes, and approve TOOL for the rest of the running session
3. No, and tell Copilot what to do differently (Esc)

Option 2 means more than it reads. Official example: after approving rm ./this-file.txt, any rm (e.g. rm -rf ./*) can run in the session. This is "allow by tool class", not per-command approval.

Unattended mode uses three switches, and deny wins:

SwitchEffect
--allow-all-toolsnever ask for any tool
--allow-toolallow specified tools
--deny-toolforbid specified tools, highest priority

Values support shell(COMMAND) (down to the shell(git push) level), write, MCP_SERVER_NAME.

Official warning: with --allow-all-tools, it has the same local access as you.

Local files and commands all count, with no per-item confirmation.

The official docs recommend using it in VMs / containers / offline environments to limit blast radius.

Trusted directories are not a hard sandbox. The official docs state plainly that the boundary

"is heuristic and GitHub does not guarantee that all files outside trusted directories will be protected"

额度:按提示词扣 premium requestsQuota: one premium request per prompt

每次提示词各扣 1 次 premium request。官方原文:

Each time you submit a prompt to Copilot CLI, your monthly quota of Copilot premium requests is reduced by one.

用非默认模型时按模型倍率放大(如 Claude Opus 4.5 为 3x)。

Free 档官方写每月最多 50 次 premium requests。付费档的内置模型(GPT-5 mini / GPT-4.1 / GPT-4o)不扣额度。

官方并宣布自 2026-06-01 起,把 Copilot 从 request-based billing 转为 usage-based billing。

档位报价见官方 plans 页(本站按本站的报价记录约定不存价格)。

Each prompt costs one premium request. Official wording:

Each time you submit a prompt to Copilot CLI, your monthly quota of Copilot premium requests is reduced by one.

Non-default models multiply by the model rate (e.g. Claude Opus 4.5 is 3x).

The Free tier officially allows up to 50 premium requests per month. Paid tiers' built-in models (GPT-5 mini / GPT-4.1 / GPT-4o) do not consume quota.

The official docs also announced that starting 2026-06-01, Copilot moves from request-based billing to usage-based billing.

Plan pricing is on the official plans page (per this site's price-recording policy, no prices are stored here).

核验说明Verification notes

confidence: partial 的依据:运行模式与审批选项已取到。MCP / 扩展面也已取到。

上下文管理与计费单位也已取到。

来源是官方 docs 与官方仓库。

剩余缺口主要是「CLI 与各档位的对应边界」。官方 about 页列 Pro/Pro+/Business/Enterprise,quickstart 页写 all Copilot plans,两处口径不一致。在本档案中记为「官方未说明」并列出所查页面,不替官方择一。

Basis for confidence: partial: execution modes and approval options were obtained. MCP / extension surfaces were also obtained.

Context management and billing units were also obtained.

Sources are official docs and the official repository.

The remaining gap is mainly "the exact mapping between the CLI and each Copilot plan". The official about page lists Pro/Pro+/Business/Enterprise, while the quickstart page says all Copilot plans — the two pages are inconsistent. This is recorded as "not stated by the official docs", listing the pages checked, without picking one on the vendor's behalf.

相关条目Related entries

  • GitHub CopilotGitHub Copilot同一产品的 IDE 形态;模型清单与云端 coding agent 见该条the IDE form of the same product; model lists and the cloud coding agent are covered there
  • Claude Code CLIClaude Code CLI同为终端编码 agent,权限审批的颗粒度可逐项对照also a terminal coding agent; permission-approval granularity can be compared item by item
  • Codex CLICodex CLI同为终端编码 agent,非交互与沙箱模型不同(Codex 默认工作区沙箱,CLI 默认逐工具审批)also a terminal coding agent, but the non-interactive and sandbox models differ (Codex defaults to a workspace sandbox; this CLI defaults to per-tool approval)
  • GitHub MCP ServerGitHub MCP Server另一层 · MCP 工具Other layer · MCP tools

本页由This page is generated from the ai-handbook 数据层生成(CC BY 4.0)。data layer (CC BY 4.0). 方法论与坐标系定义见仓库内For the methodology and coordinate definitions, see METHODOLOGY.md.