文档Docs

装好 AI 命令行工具Set up AI command-line tools

空电脑也能开始：三个系统的基础依赖、四个 AI 工具、PATH、登录和全权限设置。Start on an empty computer: dependencies, four AI tools, PATH, sign-in and full-access settings.

先选系统，再按菜单完成基础依赖和工具安装。所有命令在自己的电脑终端执行。

只想先用 OpenCode 接着安装 Agent J？装好 curl 后直接看下方 OpenCode 段的最快上手路线，不需要先装 Homebrew、Node 或 npm。Node 步骤只用于 npm 安装路线；Claude Code 和 Codex 的官方脚本也不需要 Node。

手机遥控需要工具继续执行；全权限能减少反复弹出的审批。风险：AI 可以以你的用户权限改文件和执行命令，只对可信目录和任务启用。项目或公司设置仍可能要求审批。

只想装好 Agent J 加一个 AI，不想逐个工具配置？用一条命令自己装，它会按需调用本页的脚本。

安装 Agent J · 电脑上怎么用 · agentj 命令参考

一键安装

先在下面对应系统中装好 curl。脚本询问装哪些工具，再用短超时探测官方服务连通性，不发送用户数据。能连通默认原生登录；连不通推荐 AgentsRelay，使用另一条服务入口。你可以改选，也可用 --relay / --no-relay 指定。连通不代表所在地区或账号一定可用。默认备份并合并全权限配置，开着全权限时脚本会打印警告；--no-full-access 保留默认审批。--dry-run 不下载、不写文件。macOS 和 Ubuntu/Debian/WSL2 支持自动安装。

curl -fsSL https://agentj.app/setup-agents.sh | bash

更稳妥：先下载、看内容、核对 SHA256，再执行（每条分别复制）。

curl -fsSL https://agentj.app/setup-agents.sh -o setup-agents.sh

less setup-agents.sh

shasum -a 256 setup-agents.sh

SHA256 sh: 4868818cbf2d73aeddc90cd03cc4537beac5611eb689d2e32f60ae3602abce84

bash setup-agents.sh --dry-run

bash setup-agents.sh

流程：选工具 → 选原生登录或 AgentsRelay → Relay 路线先注册、隐藏输入 key 并通过真实模型请求 → 装依赖和 CLI → 检查原生二进制并试跑离线帮助 → 添加别名 → 可选启动 Agent J 正式安装助手 → 绑定账号、配对手机。正式助手会通过真实模型调用验证并复用已有 CLI 登录，配对后直接聊天；没有可调用的受支持 CLI 时才要求新服务 key。

npm 安装失败或超时会自动重试 registry.npmmirror.com，连同平台可选依赖一起下载，不必加 --mirror；该开关仅表示优先镜像。根包与平台包的 SHA512 必须和 npmjs 官方元数据一致，下载到的压缩包也要校验。官方元数据不可用或哈希不一致就停下，不把未核实的镜像当成功。Node 压缩包也可从镜像下载，但 SHA256 只认 nodejs.org 官方 SHASUMS256.txt（仅 HTTPS、TLS 1.2 以上）；官方清单取不到或字节对不上就停下，请自行安装 Node.js 22+ 后重跑。镜像只解决下载，不改变模型服务的地区限制。脚本不从 GitHub Releases 下载 CLI 二进制；本站也不托管 Claude Code 或 Codex。

每个 CLI 都要通过 --version 和离线 --help。原生工具还会检查平台包、CPU 架构、glibc/musl 选择，试跑二进制和随包提供的 ripgrep。缺组件会报安装不完整：缺 …并给修复命令，不报成功。Gemini 还会检查并加载它的平台原生 Node 终端组件。macOS 不再按版本号拒绝，实际运行失败才说明是否需要升级系统库。

通过检查的工具会幂等添加 bash/zsh 的 cx / cc / oc / gx；Windows 助手为已检查的 WSL 工具添加 PowerShell 函数。打开新终端生效。--no-aliases 可关闭（PowerShell 为 -NoAliases）；--no-full-access 创建不带全权限参数的别名。cc 会覆盖交互终端里的 C 编译器命令；编译器可用 command cc，或把别名改成 ccx。

bash setup-agents.sh --mirror

仅配置全权限也可重跑，已有工具不会重装：

bash setup-agents.sh --cli opencode,codex --no-relay

macOS 系统

菜单：先检查系统；只装 OpenCode 可直接跳到下方 OpenCode 段。需要 npm 路线时再按 Homebrew → 依赖 → Node → PATH 继续。macOS 自带 /usr/bin/curl；找不到 curl 时直接用完整路径。

检查是否已有：

sw_vers -productVersion

uname -m

/usr/bin/curl --version

command -v brew

没有 Homebrew：官方安装器可能要求电脑密码和 Command Line Tools。

/bin/bash -c "$(/usr/bin/curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Apple Silicon（arm64）：

eval "$(/opt/homebrew/bin/brew shellenv)"

echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile

Intel（x86_64）：

eval "$(/usr/local/bin/brew shellenv)"

echo 'eval "$(/usr/local/bin/brew shellenv)"' >> ~/.zprofile

默认 zsh 写 ~/.zprofile；bash 登录终端改写 ~/.bash_profile，不要两个架构都加。

xcode-select -p

xcode-select --install

git --version

python3 --version

brew install curl git python xz

node --version

npm --version

brew install node@22

export PATH="$(brew --prefix node@22)/bin:$PATH"

echo 'export PATH="$(brew --prefix node@22)/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

Homebrew 不可用时，从 Node 官网下载与你架构相符的 LTS 安装包。

Linux 系统

菜单：系统检查 → apt 装 curl；然后可直接装下方 OpenCode。npm 路线才需要继续 Node → PATH。以下适用 Ubuntu 24.04+ / Debian 12+，其他发行版先用其包管理器装同名依赖。

uname -sm

command -v curl

git --version

python3 --version

cc --version

make --version

test -f /etc/ssl/certs/ca-certificates.crt

command -v nano

没有 curl 也能先执行 apt；sudo 要你在电脑上输入密码。容器/root 可去掉 sudo。

sudo apt-get update

sudo apt-get install -y curl git ca-certificates build-essential python3 xz-utils nano

node --version

npm --version

缺少或低于 Node 22：安装 nvm，再装 Node 22。命令把 Node/npm 放在用户目录。

curl -fsSL https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash

export NVM_DIR="$HOME/.nvm"

. "$NVM_DIR/nvm.sh"

nvm install 22

nvm alias default 22

Node 下载慢可用国内镜像重试：

NVM_NODEJS_ORG_MIRROR=https://npmmirror.com/mirrors/node nvm install 22

脚本需要 Python 3.11+；低版本先升级系统或安装发行版提供的 Python 3.11+。没有 npm 就安装 Node LTS，npm 随它一起安装；不要只装一个旧 npm。

Windows 系统

菜单：管理员 PowerShell → WSL2 → Ubuntu → 用户名/密码 → Linux 步骤 → 四个工具。Agent J 在 Windows 上只支持 WSL2 路线。需 Windows 10 build 19041+ 或 Windows 11，BIOS 虚拟化开启。

检查是否已有（管理员 PowerShell）：

wsl --status

wsl --list --verbose

wsl --install -d Ubuntu

wsl --set-default-version 2

已有 Ubuntu 但 VERSION=1：

wsl --set-version Ubuntu 2

按提示重启 Windows，打开 Ubuntu，首次建立 Linux 用户名/密码。密码输入不显示字符。

wsl -d Ubuntu

也可以下载 Windows 引导脚本，查看后运行：

Invoke-WebRequest https://agentj.app/setup-agents.ps1 -OutFile setup-agents.ps1

Get-Content ./setup-agents.ps1

Get-FileHash ./setup-agents.ps1 -Algorithm SHA256

SHA256 ps1: 0ef880556453c6a9853fc6f3e71228debec980f072b76688281c3cff58b12ab1

./setup-agents.ps1 -DryRun

./setup-agents.ps1

如果执行策略阻止脚本，不改整机策略；按上面的 wsl 命令手动安装。接下来所有 bash 命令都在 Ubuntu 窗口执行，包括 Linux Tab 的依赖和 Node 指令。工作目录建议放 ~/projects，不要放 /mnt/c。

sudo apt-get update

sudo apt-get install -y curl ca-certificates

curl -fsSL https://agentj.app/setup-agents.sh | bash

详见微软 WSL 官方指南。

安装依赖

上面的系统菜单解决缺少 curl / npm / Python / 编译工具。依赖安装失败时停在该步骤，修复网络或包管理器后重跑；脚本会给出本页对应锚点。

设置命令路径

npm 路线统一用用户目录；下面先让当前终端能找到它，以及官方安装器使用的路径。

export PATH="$HOME/.local/share/agentj-setup/npm/bin:$HOME/.local/bin:$HOME/.opencode/bin:$PATH"

zsh 写入并加载（首次执行一次）：

echo 'export PATH="$HOME/.local/share/agentj-setup/npm/bin:$HOME/.local/bin:$HOME/.opencode/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

bash：这一行必须放在 ~/.bashrc 非交互 early-return 前面；下面用 sed 插在开头，再加载。

touch ~/.bashrc

sed -i.bak '1i export PATH="$HOME/.local/share/agentj-setup/npm/bin:$HOME/.local/bin:$HOME/.opencode/bin:$PATH"' ~/.bashrc

source ~/.bashrc

macOS 的 BSD sed 不接受这条 Linux 命令；mac bash 用编辑器把同一 export 行放在 ~/.bashrc 开头，并在 ~/.bash_profile 里加载 ~/.bashrc。新开终端后再次验证。

command -v node

command -v npm

安装 OpenCode

检查是否已有：

opencode --version

command -v opencode

主路线（官方脚本；macOS / Linux / WSL，不需要 Node/npm）：

curl -fsSL https://opencode.ai/install | bash

装完提示 command not found：按自己的 shell 选择下面一组。macOS 默认 zsh；>> 会自动创建不存在的 ~/.zshrc。每条分别复制。

zsh（macOS 默认）：

echo 'export PATH="$HOME/.opencode/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

opencode --version

bash（Linux / WSL；没有 ~/.bashrc 也会自动创建）：

echo 'export PATH="$HOME/.opencode/bin:$PATH"' >> ~/.bashrc

source ~/.bashrc

opencode --version

新开终端后再验证一次。bash 的 PATH 行应放在非交互 early-return 前；macOS 若使用 bash 登录终端，还需在 ~/.bash_profile 中加载 ~/.bashrc。

npm 备用路线（先安装 Node/npm，并完成上方 PATH 设置）：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" opencode-ai

国内 npm 镜像备用：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" opencode-ai --registry=https://registry.npmmirror.com

检查 PATH，验证安装，再自己登录；密钥只在自己的终端输入。

command -v opencode

opencode --version

opencode auth login

最快上手路线（推荐给新用户）

用上方官方脚本安装 OpenCode；若找不到命令，先做紧跟脚本的 PATH 修复。
在自己的终端运行 opencode。
在 OpenCode 里输入 /connect，搜索并选择 DeepSeek，在密钥输入框粘贴自己的 API key。密钥只在自己的终端输入，不要发到聊天、席位卡或截图里。也可先在终端运行 opencode auth login，选择 DeepSeek 完成登录。
在 OpenCode 里输入 /models，选择可用的 DeepSeek 模型。
把席位卡上的安装提示词粘贴到 OpenCode 的对话中，让它接着安装 Agent J；按提示完成需要你亲自操作的步骤。

DeepSeek 官方接入指南

全权限启动：

opencode --auto

OpenCode 官方指南

当前官方 --auto 自动批准未被明确 deny 的动作；旧版报 unknown option 时先升级。永久配置见下方 permission=allow。

安装 Codex CLI

检查是否已有：

codex --version

command -v codex

主路线（官方脚本；macOS / Linux / WSL，不需要 Node/npm）：

curl -fsSL https://chatgpt.com/codex/install.sh | sh

装完提示 command not found：按自己的 shell 选择下面一组。macOS 默认 zsh；>> 会自动创建不存在的 ~/.zshrc。每条分别复制。

zsh（macOS 默认）：

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

codex --version

bash（Linux / WSL；没有 ~/.bashrc 也会自动创建）：

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc

source ~/.bashrc

codex --version

新开终端后再验证一次。bash 的 PATH 行应放在非交互 early-return 前；macOS 若使用 bash 登录终端，还需在 ~/.bash_profile 中加载 ~/.bashrc。

npm 备用路线（先安装 Node/npm，并完成上方 PATH 设置）：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @openai/codex

备用路线（brew 仅 macOS）：

brew install --cask codex

国内 npm 镜像备用：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @openai/codex --registry=https://registry.npmmirror.com

检查 PATH，验证安装，再自己登录；密钥只在自己的终端输入。

command -v codex

codex --version

codex login

全权限启动：

codex --dangerously-bypass-approvals-and-sandbox

Codex CLI 官方指南

安装 Claude Code

检查是否已有：

claude --version

command -v claude

主路线（官方脚本；macOS / Linux / WSL，不需要 Node/npm）：

curl -fsSL https://claude.ai/install.sh | bash

装完提示 command not found：按自己的 shell 选择下面一组。macOS 默认 zsh；>> 会自动创建不存在的 ~/.zshrc。每条分别复制。

zsh（macOS 默认）：

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

claude --version

bash（Linux / WSL；没有 ~/.bashrc 也会自动创建）：

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc

source ~/.bashrc

claude --version

新开终端后再验证一次。bash 的 PATH 行应放在非交互 early-return 前；macOS 若使用 bash 登录终端，还需在 ~/.bash_profile 中加载 ~/.bashrc。

npm 备用路线（先安装 Node/npm，并完成上方 PATH 设置）：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @anthropic-ai/claude-code

国内 npm 镜像备用：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @anthropic-ai/claude-code --registry=https://registry.npmmirror.com

检查 PATH，验证安装，再自己登录；密钥只在自己的终端输入。

command -v claude

claude --version

claude auth login

全权限启动：

claude --dangerously-skip-permissions

Claude Code 官方指南

安装 Gemini CLI

Google 已停用 Gemini CLI 个人登录，只能使用付费 Gemini API key。默认不推荐安装，只有明确选择 Gemini 时才安装并添加 gx 别名。

检查是否已有：

gemini --version

command -v gemini

主路线（npm；先完成 Node 和 PATH）：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @google/gemini-cli

备用路线（brew 仅 macOS）：

brew install gemini-cli

国内 npm 镜像备用：

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @google/gemini-cli --registry=https://registry.npmmirror.com

检查 PATH，验证安装，再自己登录；密钥只在自己的终端输入。

command -v gemini

gemini --version

gemini

全权限启动：

gemini --yolo --sandbox=false

Gemini CLI 官方指南

Google 已停用 Gemini CLI 个人登录，只能使用付费 Gemini API key；当前 Gemini 默认审批设置不能填 yolo，必须在启动命令里带 --yolo。Agent J 当前可接 Claude/Codex/OpenCode；Gemini 可在电脑上使用，但尚不能作为 Agent J 主 Agent。

全权限启动（三种方式）

全权限意味着 Agent 可不经询问执行命令、修改文件；只在自己的电脑和信任的目录里用。公司管理策略、显式禁止规则仍可能生效。

1. 命令参数

在终端运行所选工具的一行命令：

Codex

codex --dangerously-bypass-approvals-and-sandbox

Claude Code

claude --dangerously-skip-permissions

OpenCode

opencode --auto

Gemini

gemini --yolo --sandbox=false

OpenCode 的 --auto 自动批准原本需要询问的操作，保留显式 deny。

2. 修改配置文件

下面按工具顺序列出路径和字段。先备份，在编辑器中合并；不要把整个文件替换成示例，也不要重复添加已有键。 配置可能含 key，不要发到聊天。

工具	macOS / Linux	Windows PowerShell
Codex	`~/.codex/config.toml`	`$HOME\.codex\config.toml`
Claude Code	`~/.claude/settings.json`	`$HOME\.claude\settings.json`
OpenCode	`~/.config/opencode/opencode.json`	`$HOME\.config\opencode\opencode.json`
Gemini	`~/.gemini/settings.json`	`$HOME\.gemini\settings.json`

Codex

approval_policy = "never"
sandbox_mode = "danger-full-access"

Claude Code

{"permissions":{"defaultMode":"bypassPermissions"}}

OpenCode

{"permission":"allow"}

Gemini

{"tools":{"sandbox":false}}

Codex：两个键放在顶层、任何 [table] 之前。Claude Code：合并到用户配置的 permissions；项目配置里的 bypassPermissions 不生效。OpenCode：如果已有 opencode.jsonc，编辑它；改成 permission: "allow" 会替换原来的细分权限，只在确实希望全部放行时做。若设置了 XDG_CONFIG_HOME，使用该目录下的 opencode/opencode.json。Gemini：tools.sandbox: false 只关闭沙箱；general.defaultApprovalMode 不接受 yolo，不能单靠配置做到全部免询问，用第一种命令或下面 gx。

macOS 配置编辑

先备份，再在编辑器中手动合并上方字段；只复制所用工具的命令。WSL2 用 Linux 标签。OpenCode 若已有 jsonc 文件会优先编辑它。

Codex

file="$HOME/.codex/config.toml"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Claude Code

file="$HOME/.claude/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

OpenCode

file="${XDG_CONFIG_HOME:-$HOME/.config}/opencode/opencode.json"
test ! -f "${file}c" || file="${file}c"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Gemini

file="$HOME/.gemini/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Linux 配置编辑

先备份，再在编辑器中手动合并上方字段；只复制所用工具的命令。WSL2 用 Linux 标签。OpenCode 若已有 jsonc 文件会优先编辑它。

Codex

file="$HOME/.codex/config.toml"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Claude Code

file="$HOME/.claude/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

OpenCode

file="${XDG_CONFIG_HOME:-$HOME/.config}/opencode/opencode.json"
test ! -f "${file}c" || file="${file}c"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Gemini

file="$HOME/.gemini/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Windows 配置编辑

先备份，再在编辑器中手动合并上方字段；只复制所用工具的命令。WSL2 用 Linux 标签。OpenCode 若已有 jsonc 文件会优先编辑它。

Codex

$file = "$HOME\.codex\config.toml"
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

Claude Code

$file = "$HOME\.claude\settings.json"
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

OpenCode

$file = "$HOME\.config\opencode\opencode.json"
if ($env:XDG_CONFIG_HOME) { $file = Join-Path $env:XDG_CONFIG_HOME "opencode\opencode.json" }
if (Test-Path -LiteralPath ($file + "c")) { $file += "c" }
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

Gemini

$file = "$HOME\.gemini\settings.json"
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

回退：从对应 .backup-* / .agentj-backup 恢复原文件；新建文件只移除本次添加的字段。未备份时，把 Codex 改为 approval_policy = "on-request" / sandbox_mode = "workspace-write"，Claude 改为 defaultMode: "default"，OpenCode 改为 permission: "ask"，Gemini 移除 tools.sandbox 并去掉启动参数。退出当前 Agent，重新启动。

核验日期 2026-10-09：本机 Codex 0.159.2、Claude Code 2.1.289、OpenCode 1.18.32、Gemini 0.60.0 的 --help 及官方文档；管理策略和项目配置可能覆盖用户值。

Codex configuration · Claude Code settings · OpenCode permissions · Gemini configuration

3. 短别名

重复执行下列命令不会重复追加同一行。已有同名别名或函数时，先检查并决定是否替换；下列定义会接管同名交互命令。cx / cc / oc / gx 支持在后面继续加参数。cc 是 macOS/Linux 的系统 C 编译器名字；别名只影响交互终端，不影响普通编译脚本。不想覆盖可把命令中的 cc 改为 ccx。

回退：在对应 rc 文件或 $PROFILE 中删除本次四行；新开终端。当前 Bash/zsh 用 unalias cx cc oc gx，PowerShell 用 Remove-Item Function:cx,Function:cc,Function:oc,Function:gx。

macOS 短别名

zsh：追加到 ~/.zshrc，命令最后的 source 立即生效，也可新开终端。

Codex (cx)

rc="$HOME/.zshrc"
touch "$rc"
line="alias cx='codex --dangerously-bypass-approvals-and-sandbox'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Claude Code (cc)

rc="$HOME/.zshrc"
touch "$rc"
line="alias cc='claude --dangerously-skip-permissions'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

OpenCode (oc)

rc="$HOME/.zshrc"
touch "$rc"
line="alias oc='opencode --auto'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Gemini (gx)

rc="$HOME/.zshrc"
touch "$rc"
line="alias gx='gemini --yolo --sandbox=false'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Linux 短别名

bash：追加到 ~/.bashrc；zsh 用户切到 macOS 标签。最后 source 生效，也可新开终端。

Codex (cx)

rc="$HOME/.bashrc"
touch "$rc"
line="alias cx='codex --dangerously-bypass-approvals-and-sandbox'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Claude Code (cc)

rc="$HOME/.bashrc"
touch "$rc"
line="alias cc='claude --dangerously-skip-permissions'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

OpenCode (oc)

rc="$HOME/.bashrc"
touch "$rc"
line="alias oc='opencode --auto'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Gemini (gx)

rc="$HOME/.bashrc"
touch "$rc"
line="alias gx='gemini --yolo --sandbox=false'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Windows 短别名

PowerShell：追加函数到当前 $PROFILE，最后一行加载。也可新开同一种 PowerShell 终端；如果你的执行策略不允许加载 profile，直接用第一种启动命令，不修改公司策略。

Codex (cx)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function cx { codex --dangerously-bypass-approvals-and-sandbox @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

Claude Code (cc)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function cc { claude --dangerously-skip-permissions @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

OpenCode (oc)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function oc { opencode --auto @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

Gemini (gx)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function gx { gemini --yolo --sandbox=false @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

配置 AgentsRelay

任何 CLI 安装之前，先到 https://agentsrelay.net 注册并购买对应工具的套餐。官方引导脚本隐藏输入分组 key，实际调用模型确认可用。失败可以重贴，最多三次，也可改走原生登录。我们不发临时额度 token；默认模型由官方脚本选择，我们不写死模型名。Claude 和 OpenAI 分组 key 不同，要提供所选工具需要的分组。Gemini 只能使用付费 API key，Google 个人登录已停用。无终端时 Relay 路线会在安装 CLI 前停下，提示回自己的终端重跑。key 不要发进 AI 对话。

curl -fsSL https://agentsrelay.net/onboard/agentsrelay-setup.sh -o agentsrelay-setup.sh

sh agentsrelay-setup.sh --paste --only claude,codex,opencode --no-launch

之后新开终端，让环境变量生效。官方回滚：

python3 ~/.config/agentsrelay/agentsrelay-setup.py --restore

AgentsRelay 官方配置页

用不了 Claude Code 和 Codex 的地区：方舟或 Kimi

Anthropic 和 OpenAI 不服务香港、澳门、中国大陆等地区。在这些地方这样选：

中国大陆、从零开始：首推火山方舟 Coding Plan，先买 Lite。Lite 和 Pro 的模型、工具一样，Pro 只是 5 倍用量。
香港或海外：Kimi 会员 Plus 或更高档，用 OpenCode 或 Kimi Code 接入。在海外访问按美元收费。Go 档没有编程额度，用不了。
已经有其中一家（方舟、Kimi 或阿里百炼 Token Plan）：继续用那一家。

装好 Agent J 后，Agent 会实测你这台电脑能用什么：

agentj agent route

它向 Claude、OpenAI 和 Kimi 的服务器各问一次，不带任何 key。时区只在拿不准时参考，不会发到任何地方。Claude Code 和 Codex 用不了时，它会按上面的规则列出每家的开通步骤。Kimi 那一项会给出 Agent J 官方的邀请链接，链接后面紧跟这句话：「这是 Agent J 官方的邀请链接，您和我们各得一次抽奖」。24 小时内付费订阅才算数。订阅后在 Kimi Code 控制台新建一把 key，只粘贴进手机上的密钥卡；Agent 会把 OpenCode 切到 kimi-code-plan-cn/k3。你自己的定时任务也可以用这把会员 key（长时间连续跑的后台任务服务商可能限流，遇到了就改用 Kimi 开放平台按量付费的 key）；客服 bot 拿不到它。它也会列出备选：千问（阿里云百炼中国大陆站）、智谱 GLM / DeepSeek、OpenRouter、Kimi 开放平台按量付费。已经有其中某个服务账号的，继续用就行。kimi.com 买的会员 key 在 api.kimi.com 和 api.kimi.ai 两个地址都能用；在海外站 kimi.ai 买的会员，Agent 会改用 kimi-code-plan-global/k3。

运行要求

macOS：先试运行 opencode --version（或所选 Agent 的 --version）；能正常运行就能继续安装，不按系统版本号拦截。一键安装会实际试运行它下载的固定版 OpenCode；若实际出现 dyld / _ubrk_clone 或无法运行，才提示失败原因。PATH 或 npm 重装不能修复系统库，请升级系统、换电脑，或用能正常运行的其他 Agent 按手动路线安装。Intel Mac 使用 x86_64 工具与 /usr/local Homebrew，Apple Silicon 使用 arm64 与 /opt/homebrew；不要混用架构。Gemini 当前官方列 macOS 15+。物理 Intel Mac/Windows 安装尚需对应机器验证。

中国大陆：Claude Code / Codex 官方服务不可用，npm 镜像只帮你下载，不能解锁服务。选择 OpenCode + 国产模型，或使用有对应套餐/模型的 AgentsRelay；需自己核对服务提供方和地区条款。Windows 只走 WSL2。

command not found：回 PATH 步骤，确认 shell 文件和新开终端。EACCES：使用本页用户目录 npm prefix，不 sudo npm。登录失败：核对账号、地区和自己的代理，别把 key 贴出来。

Agent 隔离与首次设置

Linux 上 Agent J 用 bubblewrap 把手机上的 Agent 关在隔离里运行；socat 是 Claude Code 自带沙箱要用的。一键安装会用 apt-get 装上这两个；装不了时会打印要你自己运行的命令：sudo apt-get install -y bubblewrap socat（Fedora 用 dnf，Arch 用 pacman）。不装的话，每次启动 Agent 手机上都要用通行密钥批准「本次不隔离运行」。

选了 Claude Code 时，脚本还会预先完成它的首次设置（引导、主题、危险模式确认），后台会话不会停在没人看的终端确认上；只补缺的，不改你设过的。用 root 运行时不写全权限（bypassPermissions）：Claude Code 不允许 root 这样做，会直接退出。

安装 Agent J

结尾会问现在安装 Agent J 吗？[Y/n]。选 Y 就调用现有正式安装助手，需要账户页的安装码（AJI-…），在本机隐藏输入。--with-agentj 选择这一步，--no-agentj 跳过；账号确认和手机配对仍要你本人操作。无终端时只打印下一步。没检测到登录或 key 时，会先提示：Agent J 能装能配对，但要登录 AI 工具或填 key 后才能真正聊天。检测到凭据不等于确认它仍有效。

问 Installation code (AJI-…) 时，在账户页点「复制安装码」，只粘贴这串 AJI- 开头的码。不小心粘贴了整段提示词或安装命令也没关系，脚本会从中取出第一张 AJI 码；全角字符、前后空格和换行都会自动处理。

正式助手会实际调用已安装的 Claude Code、Codex 和 OpenCode，验证能回复后复用本机登录；多个可用时由你选择。手机配对后，发 Hello 就能开始聊天。都不能调用时，才继续填写正式服务地址、模型和手机密钥卡。AgentsRelay 已配好的凭据也能由后台服务读取，无需新开终端。Gemini 可以由本页安装，但当前还不能作为 Agent J 的主 Agent。

Please download https://agentj.app/install.md with curl, read the whole file with your file-reading tool, then follow it step by step, re-reading each section before you do it. Tell me when you need me to do something.

安装流程 · install.md · 命令参考

让电脑保持唤醒：命令、系统设置和可发给 Agent 的请求。

仅有 Gemini 时，安装助手会提示：Agent J 主 Agent 暂不支持 Gemini，请再登录 Codex / Claude Code / OpenCode 之一或用 AgentsRelay。setup-agents 仍可安装 Gemini。

安装 Kimi Code CLI

Kimi Code CLI（kimi，月之暗面，MIT 开源）是 Agent J 可以直接用的 AI 编程工具。官方脚本不需要 Node.js：

curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash

Windows（PowerShell）：irm https://code.kimi.com/kimi-code/install.ps1 | iex（先装 Git for Windows）。也可以用 npm：npm install -g @moonshot-ai/kimi-code（在中国大陆加 --registry=https://registry.npmmirror.com）。装好用 kimi --version 检查，Agent J 需要 2.2.0 或更新。

在你自己的终端登录：国内站 kimi.com 用 kimi login --region mainland-cn，海外站 kimi.ai 用 kimi login --region global（在 kimi.ai 买的会员）。它会给出一个链接和一次性设备码，在任何登录了 Kimi 的设备上确认就行。然后运行 agentj agent switch kimi。

隔离：Kimi 没有自带沙箱，它的 hook 出错时会放行，所以 Agent J 让它和别的工具一样在隔离里运行，固定在手动审批模式，每个权限请求都发到你手机上。

阿里百炼 Token Plan（中国大陆）

给编程工具用的包月套餐，不按 API 用量计费。Token Plan 个人版分 Lite、Essential、Standard、Pro 四档，按 Credits 计量（每月 11,500 / 25,500 / 45,000 / 180,000；Standard 起可以同时跑 3–4 个 Agent）。每档月费让 Agent 用 agentj agent route 列给你看，以阿里云订阅页为准。千问 3.8 Max 和 3.8 Flash 能看图，截图和浏览器自动化都能用；Agent J 默认用 qwen3.8-max，想省额度可以换 qwen3.8-flash。

买之前先看清条款，以阿里云订阅页为准。你自己的定时任务和后台 Agent 可以用 Token Plan 的 key。长时间连续跑的后台任务，服务商可能限流，遇到了就把这些任务改用按量付费的 key。订阅期间，你发给模型的内容和模型生成的内容会被阿里云用于服务改进和模型优化，涉及隐私的数据要谨慎。套餐只限订阅人本人，不能共享。客服 bot 是在服务别人，Agent J 不会把 Token Plan 的 key 交给 bot，要给 bot 另填一把按量付费的 key（bot 设置页里会告诉你填在哪）。

在 https://bailian.console.aliyun.com/cn-beijing/subscription/overview 订阅，再到控制台新建套餐专属 API Key。然后让 Agent 帮你配：它运行 agentj provider bailian，把百炼写进 OpenCode 的设置，你原来的设置一项都不动。接着手机上会弹出密钥卡，key 只粘贴在那里。Agent J 会用百炼只读的模型列表接口核对这把 key，不花 Credits。最后 Agent 把 OpenCode 切到 bailian-token-plan-personal/qwen3.8-max。

Coding Plan（key 以 sk-sp- 开头）也能用（agentj provider bailian --plan coding-plan），但它没有 Max 模型，key 也没法事先核对。你自己的定时任务可以用 Coding Plan 的 key（长时间连续跑的后台任务服务商可能限流，遇到了就改用按量付费的 key）；客服 bot 拿不到它，要另填按量付费的 key。

千问 Qwen Code

Qwen Code（qwen，阿里，Apache-2.0 开源）是阿里自己的编程 Agent，Agent J 可以直接用它，用的还是同一把百炼 key。用 npm 安装（需要 Node.js 20 或更新；在中国大陆加 --registry=https://registry.npmmirror.com）：

npm install -g @qwen-code/qwen-code@latest

装好用 qwen --version 检查，Agent J 需要 0.25 或更新。它的免费登录 2026-04-15 已经停了，所以 key 就是它的登录：用上面那张手机密钥卡（BAILIAN_TOKEN_PLAN_API_KEY）。Agent J 只在内存里把 key 交给 Qwen Code，不往 ~/.qwen 写任何东西。如果你自己已经配好了 Qwen Code（比如用过 bl config agent --agent qwen-code），就用你自己的设置。然后运行 agentj agent switch qwen。

隔离和审批：Agent J 让 Qwen Code 和别的工具一样在隔离里运行。Qwen Code 默认的模式会自己批准一部分操作，Agent J 每次都把它切回「每次都问」，每个请求都发到你手机上，只批这一次。手机上的 /model 会列出套餐里的模型，名字后面带“看图”两个字的能读图。

火山方舟 Coding Plan（中国大陆）

适合谁：在中国大陆、还没有任何套餐的人，我们首推它；已经开通的人也直接用它。一个套餐就能用字节豆包、Kimi、智谱 GLM、DeepSeek 和 MiniMax 的模型。先买 Lite：Lite 和 Pro 的模型、工具一样，Pro 只是 5 倍用量，Lite 不够用再升级。agentj agent route 会显示标价，以下单页为准。

哪些模型能看图（2026-10-15 实测）：doubao-seed-2.1-pro（默认）、kimi-k3（扣额度快，方舟建议 Pro 用，Lite 也能选）、kimi-k2.7-code、glm-5-3-flash、minimax-m3、deepseek-v4-1-flash 能看截图；glm-5.3、deepseek-v4-pro、deepseek-v4-flash 是纯文本。Agent J 会把这些写进 OpenCode 的设置，同时调高输出上限（GLM-5.3 的思考会用光 OpenCode 默认的 32k）。Coding Plan 不含生图模型。

这是包月套餐，条款以方舟控制台为准。你自己的定时任务和后台 Agent 可以用 Coding Plan 的 key。长时间连续跑的后台任务，服务商可能限流，遇到了就把这些任务改用按量付费的方舟 key。客服 bot 是在服务别人，Agent J 不会把 Coding Plan 的 key 交给 bot，要给 bot 另填一把按量付费的 key（bot 设置页里会告诉你填在哪）。

在 https://www.volcengine.com/activity/codingplan 用手机号注册火山引擎、完成实名，订阅 Coding Plan。再到方舟控制台（https://console.volcengine.com/ark/）新建一把 API Key，然后让 Agent 帮你配：用 OpenCode 时它运行 agentj provider ark，把套餐里的模型全部写进 OpenCode 的设置，你原来的设置一项都不动。接着手机上会弹出密钥卡，key 只粘贴在那里。方舟没有只读的核对接口，所以第一条回复就是核对。最后 Agent 把 OpenCode 切到 ark-coding-plan/doubao-seed-2.1-pro。

想用 Claude Code？让 Agent 改为给 Claude Code 配方舟，它按 Agent J 打印的三步做，key 还是走手机密钥卡。Claude Code 的 Haiku、Sonnet、Opus 默认模型和子 agent 模型都会指向同一个方舟模型，后台任务就不会去向方舟要 Claude 的模型名。

DeepSeek 编程工具 dsh

dsh（DeepSeek Harness，DeepSeek 出品，MIT 开源）是 DeepSeek 自己的编程 Agent，现在还是开发者预览版。Olares 盒子自带的 Agent「Lares」里跑的也是它，所以在 Olares 上，它和盒子自己的 Router 正好配成一对。用 npm 安装（需要 Node.js 22.19 或更新，大约占 500 MB）：

npm install -g @deepseek-ai/dsh@latest

装好用 dsh --version 检查，Agent J 需要 0.2.0-rc.2 或更新。然后选模型从哪儿来：

你自己的模型服务（Olares Router、LiteLLM、Ollama、llama.cpp、vLLM，或任何 OpenAI 兼容服务）：让 Agent 运行 agentj agent dsh --dir <目录> --base-url https://<服务>/v1 --model <模型 id>。Agent J 把地址和模型存进 ~/.dsh/agentj-route.yml，不往里写 key。明文 http:// 只能用于本机或家里的局域网。服务要 key 的话，手机上会弹出密钥卡（DSH_ROUTER_API_KEY），key 只粘贴在那里。存好后再发一次消息就行，不用重启。
DeepSeek 官方 API（按量付费）：同样走手机密钥卡，存 DEEPSEEK_API_KEY。
你自己已经配好了 dsh（比如在 dsh web 的模型页里配过）：就用你自己的设置。

隔离和审批：Agent J 让 dsh 以只读方式启动，每次都问。改文件、跑不是只读的命令，每一个都变成手机卡片，只批这一次。dsh 自己的使用数据上报已经关掉。dsh 暂时还不能看图；DeepSeek 也说，预览版以后可能还会大改。

Choose your system, then follow the menu through dependencies and tools. Run commands in your own computer’s terminal.

Want to use OpenCode to finish installing Agent J? Once curl is available, go straight to the OpenCode section and its fastest start. You do not need Homebrew, Node or npm first. Node steps are for npm installation routes; the official Claude Code and Codex scripts also work without Node.

Full access reduces repeated approval dialogs during phone control. Risk: the AI can modify files and run commands with your user permissions; enable it only for trusted folders and tasks. Project or company policies may still require approval.

Just want Agent J plus one AI, without setting up each tool? Use one command; it runs this page's script when it needs it.

Install Agent J · On your computer · agentj command reference

One-command setup

First install curl below. The setup asks which tools to install, checks official service connectivity with short requests that send no user data, and recommends a route. Reachable: native login. Unreachable: AgentsRelay, which provides an alternative service endpoint. You can change the recommendation with --relay / --no-relay. Connectivity does not prove regional eligibility or account access. Full-access settings are backed up and merged, and the script prints a warning while full access is on; --no-full-access keeps approval defaults. --dry-run downloads nothing and writes no files. Automatic setup supports macOS and Ubuntu/Debian/WSL2.

curl -fsSL https://agentj.app/setup-agents.sh | bash

To review first: download, inspect, compare SHA256, then execute. Copy each command separately.

curl -fsSL https://agentj.app/setup-agents.sh -o setup-agents.sh

less setup-agents.sh

shasum -a 256 setup-agents.sh

SHA256 sh: 4868818cbf2d73aeddc90cd03cc4537beac5611eb689d2e32f60ae3602abce84

bash setup-agents.sh --dry-run

bash setup-agents.sh

Flow: choose tools → choose native login or AgentsRelay → on Relay, register and paste keys with hidden input, pass a real model request → install dependencies/CLIs → verify native binaries and offline help → add aliases → optionally run the official Agent J installation assistant → account linking and phone pairing. The official assistant verifies existing CLI sign-in with a real model call and reuses it for immediate chat after pairing. It asks for a new formal provider key only when no supported CLI can call.

npm failure or timeout automatically retries registry.npmmirror.com, including platform optional dependencies. No --mirror is needed; it selects the mirror first. Mirror root and platform-package SHA512 integrity must equal the official npmjs metadata, and downloaded tarball bytes must match. If official metadata is unavailable or hashes disagree, setup stops rather than trusting an unverified mirror. Node tarballs may also come from the mirror, but the SHA256 always comes from the official nodejs.org SHASUMS256.txt (HTTPS only, TLS 1.2+); if that list is unreachable or the bytes differ, setup stops and asks you to install Node.js 22+ yourself. Mirrors change downloads, not model service access. CLI binaries are never downloaded from GitHub Releases by this script; neither Claude Code nor Codex is hosted on our site.

Each installed CLI must pass --version and offline --help. Native tools additionally check platform packages, CPU architecture and glibc/musl selection, execute their binaries and bundled ripgrep where provided. Missing components produce “安装不完整：缺 …” and a repair command, never success. Gemini also checks and loads its platform-native Node terminal addon. macOS has no version-number rejection; actual runtime errors explain when system libraries need an upgrade.

Successful tools get idempotent cx / cc / oc / gx aliases in bash/zsh; the Windows helper adds PowerShell functions for the verified WSL tools. Open a new terminal. --no-aliases disables this (-NoAliases in PowerShell); --no-full-access creates plain aliases. cc shadows the interactive C compiler command; use command cc or rename the alias to ccx.

bash setup-agents.sh --mirror

Rerun for full-access configuration; existing tools are not reinstalled:

bash setup-agents.sh --cli opencode,codex --no-relay

macOS

Menu: check the system first; for OpenCode alone, go straight to its section below. For npm routes, continue through Homebrew → dependencies → Node → PATH. macOS includes /usr/bin/curl; use its full path if curl is missing.

Check what is already installed:

sw_vers -productVersion

uname -m

/usr/bin/curl --version

command -v brew

No Homebrew: its official installer may request your computer password and Command Line Tools.

/bin/bash -c "$(/usr/bin/curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Apple Silicon (arm64):

eval "$(/opt/homebrew/bin/brew shellenv)"

echo 'eval "$(/opt/homebrew/bin/brew shellenv)"' >> ~/.zprofile

Intel (x86_64):

eval "$(/usr/local/bin/brew shellenv)"

echo 'eval "$(/usr/local/bin/brew shellenv)"' >> ~/.zprofile

Use ~/.zprofile for default zsh, or ~/.bash_profile for a bash login terminal. Add only your architecture’s command.

xcode-select -p

xcode-select --install

git --version

python3 --version

brew install curl git python xz

node --version

npm --version

brew install node@22

export PATH="$(brew --prefix node@22)/bin:$PATH"

echo 'export PATH="$(brew --prefix node@22)/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

If Homebrew is unavailable, download the LTS installer for your architecture from Node.js.

Linux

Menu: system check → install curl with apt → OpenCode below. Continue through Node → PATH only for npm routes. These commands target Ubuntu 24.04+ / Debian 12+; on other distributions use their package manager for equivalent prerequisites.

uname -sm

command -v curl

git --version

python3 --version

cc --version

make --version

test -f /etc/ssl/certs/ca-certificates.crt

command -v nano

apt works before curl is installed. Enter sudo’s password on your computer; omit sudo when already root.

sudo apt-get update

sudo apt-get install -y curl git ca-certificates build-essential python3 xz-utils nano

node --version

npm --version

Missing Node/npm, or Node older than 22: install nvm, then Node 22 in your user directory.

curl -fsSL https://raw.githubusercontent.com/nvm-sh/nvm/v0.40.3/install.sh | bash

export NVM_DIR="$HOME/.nvm"

. "$NVM_DIR/nvm.sh"

nvm install 22

nvm alias default 22

For slow Node downloads, retry with this mirror:

NVM_NODEJS_ORG_MIRROR=https://npmmirror.com/mirrors/node nvm install 22

The setup script requires Python 3.11+. Upgrade the OS or install a distribution-supported Python 3.11+ if needed. Installing Node LTS also installs npm; do not install an old standalone npm.

Windows

Menu: administrator PowerShell → WSL2 → Ubuntu → Linux username/password → Linux steps → four tools. Agent J on Windows uses WSL2 only. Requires Windows 10 build 19041+ or Windows 11 with BIOS virtualization enabled.

Check what is installed (administrator PowerShell):

wsl --status

wsl --list --verbose

wsl --install -d Ubuntu

wsl --set-default-version 2

If Ubuntu exists but shows VERSION=1:

wsl --set-version Ubuntu 2

Restart Windows if asked, open Ubuntu and create a Linux username/password on first launch. Password input is hidden.

wsl -d Ubuntu

Alternatively download and review the Windows helper:

Invoke-WebRequest https://agentj.app/setup-agents.ps1 -OutFile setup-agents.ps1

Get-Content ./setup-agents.ps1

Get-FileHash ./setup-agents.ps1 -Algorithm SHA256

SHA256 ps1: 0ef880556453c6a9853fc6f3e71228debec980f072b76688281c3cff58b12ab1

./setup-agents.ps1 -DryRun

./setup-agents.ps1

If execution policy blocks the helper, use the manual wsl commands above. Run every subsequent bash command inside Ubuntu, including the Linux tab’s dependencies and Node steps. Keep projects under ~/projects, rather than /mnt/c.

sudo apt-get update

sudo apt-get install -y curl ca-certificates

curl -fsSL https://agentj.app/setup-agents.sh | bash

See Microsoft’s WSL installation guide.

Dependencies

The system menus above cover missing curl, npm, Python and build tools. Fix network/package-manager failures at that step, then rerun; the script links back to the relevant section.

Command path

The npm route uses a user-owned prefix. First add it and the official installers’ directories to this terminal’s PATH.

export PATH="$HOME/.local/share/agentj-setup/npm/bin:$HOME/.local/bin:$HOME/.opencode/bin:$PATH"

zsh: save and load once:

echo 'export PATH="$HOME/.local/share/agentj-setup/npm/bin:$HOME/.local/bin:$HOME/.opencode/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

bash: place the line before any noninteractive early-return in ~/.bashrc. These commands prepend it, then load it.

touch ~/.bashrc

sed -i.bak '1i export PATH="$HOME/.local/share/agentj-setup/npm/bin:$HOME/.local/bin:$HOME/.opencode/bin:$PATH"' ~/.bashrc

source ~/.bashrc

macOS BSD sed uses different syntax; for mac bash, use an editor to put the same export line first in ~/.bashrc and load ~/.bashrc from ~/.bash_profile. Verify again in a new terminal.

command -v node

command -v npm

OpenCode

Check what is already installed:

opencode --version

command -v opencode

Primary route (official script; macOS / Linux / WSL, no Node/npm required):

curl -fsSL https://opencode.ai/install | bash

If installation ends with command not found, use the commands for your shell. macOS defaults to zsh; >> creates ~/.zshrc if it does not exist. Copy each command separately.

zsh (macOS default):

echo 'export PATH="$HOME/.opencode/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

opencode --version

bash (Linux / WSL; creates ~/.bashrc if missing):

echo 'export PATH="$HOME/.opencode/bin:$PATH"' >> ~/.bashrc

source ~/.bashrc

opencode --version

Verify once more in a new terminal. Put the bash PATH line before any noninteractive early-return; for a macOS bash login terminal, also load ~/.bashrc from ~/.bash_profile.

Alternative npm route (install Node/npm and complete the PATH section above first):

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" opencode-ai

Alternative npm download mirror for mainland China:

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" opencode-ai --registry=https://registry.npmmirror.com

Check PATH, verify the installation, then sign in yourself. Enter keys only in your own terminal.

command -v opencode

opencode --version

opencode auth login

Fastest start (recommended for new users)

Install OpenCode with the official script above; if the command is missing, use the PATH repair immediately below it.
Run opencode in your own terminal.
In OpenCode, enter /connect, search for DeepSeek and select it, then paste your own API key into the key prompt. Enter keys only in your own terminal; do not share them in chat, seat cards or screenshots. Alternatively, run opencode auth login in the terminal and select DeepSeek to sign in first.
Enter /models in OpenCode and select an available DeepSeek model.
Paste the installation prompt from your seat card into the OpenCode conversation and let it continue installing Agent J. Follow any steps it asks you to perform yourself.

DeepSeek connection guide

Full-access launch:

opencode --auto

OpenCode official guide

Current --auto approves actions that are not explicitly denied. Upgrade if an older version reports unknown option. The permanent permission=allow configuration is below.

Codex CLI

Check what is already installed:

codex --version

command -v codex

Primary route (official script; macOS / Linux / WSL, no Node/npm required):

curl -fsSL https://chatgpt.com/codex/install.sh | sh

If installation ends with command not found, use the commands for your shell. macOS defaults to zsh; >> creates ~/.zshrc if it does not exist. Copy each command separately.

zsh (macOS default):

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

codex --version

bash (Linux / WSL; creates ~/.bashrc if missing):

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc

source ~/.bashrc

codex --version

Verify once more in a new terminal. Put the bash PATH line before any noninteractive early-return; for a macOS bash login terminal, also load ~/.bashrc from ~/.bash_profile.

Alternative npm route (install Node/npm and complete the PATH section above first):

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @openai/codex

Alternative route (brew commands are macOS only):

brew install --cask codex

Alternative npm download mirror for mainland China:

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @openai/codex --registry=https://registry.npmmirror.com

Check PATH, verify the installation, then sign in yourself. Enter keys only in your own terminal.

command -v codex

codex --version

codex login

Full-access launch:

codex --dangerously-bypass-approvals-and-sandbox

Codex CLI official guide

Claude Code

Check what is already installed:

claude --version

command -v claude

Primary route (official script; macOS / Linux / WSL, no Node/npm required):

curl -fsSL https://claude.ai/install.sh | bash

If installation ends with command not found, use the commands for your shell. macOS defaults to zsh; >> creates ~/.zshrc if it does not exist. Copy each command separately.

zsh (macOS default):

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.zshrc

source ~/.zshrc

claude --version

bash (Linux / WSL; creates ~/.bashrc if missing):

echo 'export PATH="$HOME/.local/bin:$PATH"' >> ~/.bashrc

source ~/.bashrc

claude --version

Verify once more in a new terminal. Put the bash PATH line before any noninteractive early-return; for a macOS bash login terminal, also load ~/.bashrc from ~/.bash_profile.

Alternative npm route (install Node/npm and complete the PATH section above first):

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @anthropic-ai/claude-code

Alternative npm download mirror for mainland China:

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @anthropic-ai/claude-code --registry=https://registry.npmmirror.com

Check PATH, verify the installation, then sign in yourself. Enter keys only in your own terminal.

command -v claude

claude --version

claude auth login

Full-access launch:

claude --dangerously-skip-permissions

Claude Code official guide

Gemini CLI

Check what is already installed:

gemini --version

command -v gemini

Primary route (npm; complete Node and PATH first):

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @google/gemini-cli

Alternative route (brew commands are macOS only):

brew install gemini-cli

Alternative npm download mirror for mainland China:

npm install --global --prefix "$HOME/.local/share/agentj-setup/npm" @google/gemini-cli --registry=https://registry.npmmirror.com

Check PATH, verify the installation, then sign in yourself. Enter keys only in your own terminal.

command -v gemini

gemini --version

gemini

Full-access launch:

gemini --yolo --sandbox=false

Gemini CLI official guide

Google has discontinued personal Gemini CLI login. Use a paid Gemini API key, entered only in your own terminal. Gemini is optional: select it explicitly; setup does not recommend it by default. The defaultApprovalMode setting does not accept yolo; keep --yolo in the launch command. Agent J currently connects to Claude/Codex/OpenCode; Gemini works on the computer but is not yet an Agent J host adapter.

Full-access launch (three ways)

Full access lets the Agent run commands and modify files without asking. Use it only on your own computer and in trusted directories. Managed policies and explicit deny rules can still apply.

1. Command flags

Run the line for your chosen tool in your terminal:

Codex

codex --dangerously-bypass-approvals-and-sandbox

Claude Code

claude --dangerously-skip-permissions

OpenCode

opencode --auto

Gemini

gemini --yolo --sandbox=false

OpenCode --auto approves requests that would ask, while preserving explicit deny rules.

2. Configuration files

Paths and fields follow the tool order below. Back up first, then merge in an editor; do not replace the whole file with an example or duplicate existing keys. Files may contain credentials; never paste them into chat.

Tool	macOS / Linux	Windows PowerShell
Codex	`~/.codex/config.toml`	`$HOME\.codex\config.toml`
Claude Code	`~/.claude/settings.json`	`$HOME\.claude\settings.json`
OpenCode	`~/.config/opencode/opencode.json`	`$HOME\.config\opencode\opencode.json`
Gemini	`~/.gemini/settings.json`	`$HOME\.gemini\settings.json`

Codex

approval_policy = "never"
sandbox_mode = "danger-full-access"

Claude Code

{"permissions":{"defaultMode":"bypassPermissions"}}

OpenCode

{"permission":"allow"}

Gemini

{"tools":{"sandbox":false}}

Codex: put both keys at the root, before any [table]. Claude Code: merge into permissions in user settings; project settings cannot enable bypassPermissions. OpenCode: edit existing opencode.jsonc if present; setting permission: "allow" replaces granular permissions, so choose it only when you intend to allow everything. With XDG_CONFIG_HOME, use opencode/opencode.json under that directory. Gemini: tools.sandbox: false only disables the sandbox; general.defaultApprovalMode does not accept yolo. Configuration alone cannot enable full auto-approval; use the command flag or gx below.

macOS configuration

Back up, then manually merge the fields above in the editor. Copy only your tool’s command. In WSL2 use the Linux tab. An existing OpenCode jsonc file is preferred.

Codex

file="$HOME/.codex/config.toml"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Claude Code

file="$HOME/.claude/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

OpenCode

file="${XDG_CONFIG_HOME:-$HOME/.config}/opencode/opencode.json"
test ! -f "${file}c" || file="${file}c"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Gemini

file="$HOME/.gemini/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Linux configuration

Back up, then manually merge the fields above in the editor. Copy only your tool’s command. In WSL2 use the Linux tab. An existing OpenCode jsonc file is preferred.

Codex

file="$HOME/.codex/config.toml"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Claude Code

file="$HOME/.claude/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

OpenCode

file="${XDG_CONFIG_HOME:-$HOME/.config}/opencode/opencode.json"
test ! -f "${file}c" || file="${file}c"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Gemini

file="$HOME/.gemini/settings.json"
mkdir -p "$(dirname "$file")"
if test -f "$file"; then cp -p "$file" "$file.backup-$(date +%Y%m%d-%H%M%S)"; fi
nano "$file"

Windows configuration

Back up, then manually merge the fields above in the editor. Copy only your tool’s command. In WSL2 use the Linux tab. An existing OpenCode jsonc file is preferred.

Codex

$file = "$HOME\.codex\config.toml"
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

Claude Code

$file = "$HOME\.claude\settings.json"
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

OpenCode

$file = "$HOME\.config\opencode\opencode.json"
if ($env:XDG_CONFIG_HOME) { $file = Join-Path $env:XDG_CONFIG_HOME "opencode\opencode.json" }
if (Test-Path -LiteralPath ($file + "c")) { $file += "c" }
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

Gemini

$file = "$HOME\.gemini\settings.json"
New-Item -ItemType Directory -Force (Split-Path -Parent $file) | Out-Null
if (Test-Path -LiteralPath $file) { Copy-Item -LiteralPath $file -Destination ($file + ".backup-" + (Get-Date -Format yyyyMMdd-HHmmss)) }
notepad $file

Rollback: restore the matching .backup-* / .agentj-backup file. For a new file remove only the fields you added. Without a backup, use Codex approval_policy = "on-request" / sandbox_mode = "workspace-write", Claude defaultMode: "default", OpenCode permission: "ask", and remove Gemini tools.sandbox and launch flags. Exit the Agent and start a fresh session.

Checked 2026-10-09 against local Codex 0.159.2, Claude Code 2.1.289, OpenCode 1.18.32 and Gemini 0.60.0 --help and official documentation. Managed policy and project configuration may override user settings.

Codex configuration · Claude Code settings · OpenCode permissions · Gemini configuration

3. Short aliases

Add gx only if you explicitly choose Gemini and have a paid Gemini API key.

Repeating the command does not append the same line twice. Check existing aliases/functions with these names before replacing them; these definitions take over matching interactive commands. Add arguments after cx / cc / oc / gx as usual. cc is the system C compiler name on macOS/Linux: the alias affects interactive terminals, not ordinary build scripts. Use ccx instead by changing cc in the command if you prefer.

Rollback: delete the four added lines from the rc file or $PROFILE and open a new terminal. In the current Bash/zsh shell run unalias cx cc oc gx; in PowerShell run Remove-Item Function:cx,Function:cc,Function:oc,Function:gx.

macOS aliases

zsh: append to ~/.zshrc. The final source applies it now, or open a new terminal.

Codex (cx)

rc="$HOME/.zshrc"
touch "$rc"
line="alias cx='codex --dangerously-bypass-approvals-and-sandbox'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Claude Code (cc)

rc="$HOME/.zshrc"
touch "$rc"
line="alias cc='claude --dangerously-skip-permissions'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

OpenCode (oc)

rc="$HOME/.zshrc"
touch "$rc"
line="alias oc='opencode --auto'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Gemini (gx)

rc="$HOME/.zshrc"
touch "$rc"
line="alias gx='gemini --yolo --sandbox=false'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Linux aliases

bash: append to ~/.bashrc; zsh users use the macOS tab. The final source applies it now, or open a new terminal.

Codex (cx)

rc="$HOME/.bashrc"
touch "$rc"
line="alias cx='codex --dangerously-bypass-approvals-and-sandbox'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Claude Code (cc)

rc="$HOME/.bashrc"
touch "$rc"
line="alias cc='claude --dangerously-skip-permissions'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

OpenCode (oc)

rc="$HOME/.bashrc"
touch "$rc"
line="alias oc='opencode --auto'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Gemini (gx)

rc="$HOME/.bashrc"
touch "$rc"
line="alias gx='gemini --yolo --sandbox=false'"
grep -Fqx -- "$line" "$rc" || printf '\n%s\n' "$line" >> "$rc"
source "$rc"

Windows aliases

PowerShell: append functions to the current $PROFILE, then load it with the final line. Or open a new terminal of the same PowerShell edition. If policy blocks profile loading, use the direct launch commands; do not change company policy.

Codex (cx)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function cx { codex --dangerously-bypass-approvals-and-sandbox @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

Claude Code (cc)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function cc { claude --dangerously-skip-permissions @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

OpenCode (oc)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function oc { opencode --auto @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

Gemini (gx)

$dir = Split-Path -Parent $PROFILE
New-Item -ItemType Directory -Force $dir | Out-Null
if (!(Test-Path -LiteralPath $PROFILE)) { New-Item -ItemType File $PROFILE | Out-Null }
$line = 'function gx { gemini --yolo --sandbox=false @args }'
if (@(Get-Content -LiteralPath $PROFILE) -cnotcontains $line) { Add-Content -LiteralPath $PROFILE -Value ("`n" + $line) }
. $PROFILE

AgentsRelay

Before installing any CLI, visit https://agentsrelay.net, register and buy a plan for your tools. The official helper reads group keys with hidden input and performs real model calls. A failed key can be pasted again, up to three attempts, or you can switch to native login. We issue no temporary allowance token. The helper owns model defaults; the setup script hardcodes no model name. Claude and OpenAI group keys differ: provide the groups needed by your selected tools. Gemini requires a paid Gemini API key; personal Google login has been discontinued. Without a terminal, Relay stops before installing tools and asks you to rerun interactively. Never send keys in AI chat.

curl -fsSL https://agentsrelay.net/onboard/agentsrelay-setup.sh -o agentsrelay-setup.sh

sh agentsrelay-setup.sh --paste --only claude,codex,opencode --no-launch

Open a new terminal afterward to load the environment. Official rollback:

python3 ~/.config/agentsrelay/agentsrelay-setup.py --restore

AgentsRelay official onboarding

Where Claude Code and Codex are not offered: Ark or Kimi

Anthropic and OpenAI do not serve Hong Kong, Macau, mainland China and some other places. There, pick like this:

Mainland China, starting from zero: a Volcano Ark Coding Plan. Buy Lite first: it has the same models and tools as Pro, and Pro only gives 5x the quota.
Hong Kong or overseas: a Kimi membership, Plus or higher, used through OpenCode or Kimi Code. Outside the mainland it is billed in USD. The Go plan has no coding quota and does not work.
Already paying for one of them (Ark, Kimi or an Alibaba Cloud Bailian Token Plan): keep it.

After installing Agent J, your Agent checks what really works from your computer:

agentj agent route

It asks the Claude, OpenAI and Kimi servers once each, without any key. Your time zone only breaks a tie and is never sent anywhere. When Claude Code and Codex are not available, it prints the rule above with the steps for each plan. For Kimi it shows Agent J's official invite link with this sentence right next to it: "This is Agent J's official invite link: you and we each get one prize draw". The draw counts when you subscribe within 24 hours. Then create a key in the Kimi Code console and paste it only into the key card on your phone; your Agent switches OpenCode to kimi-code-plan-cn/k3. Your own scheduled tasks may use this membership key too (the provider may rate-limit long, continuous background work; switch those to a Kimi pay-as-you-go key if it does); a support bot never gets it. It also lists alternatives: Qwen (Alibaba Cloud Bailian, mainland China site), Zhipu GLM / DeepSeek, OpenRouter and Kimi's pay-as-you-go platform. If you already have an account with one of them, keep using it. A kimi.com membership key works on both api.kimi.com and api.kimi.ai; for a membership bought on kimi.ai, your Agent uses kimi-code-plan-global/k3.

System requirements

macOS: first run opencode --version (or your selected agent’s --version). If it works, continue installing; there is no operating-system version gate. The one-line installer probes its downloaded, pinned OpenCode binary and reports actual dyld / _ubrk_clone or execution failures. PATH changes or npm reinstalls cannot repair system libraries; upgrade macOS, use another computer, or choose another working agent with the manual installation guide. Intel Mac uses x86_64 tools and /usr/local Homebrew; Apple Silicon uses arm64 and /opt/homebrew. Gemini currently lists macOS 15+. Physical Intel Mac and Windows installation still requires verification on those machines.

Mainland China: official Claude Code / Codex services are unavailable; npm mirrors help downloads, not service access. Use OpenCode with a domestic model, or an appropriate AgentsRelay plan/model, checking the provider’s terms and regional requirements. Windows uses WSL2.

command not found: revisit PATH, check the shell startup file and open a fresh terminal. EACCES: use the user-owned npm prefix above, rather than sudo npm. Sign-in failure: check account, region and your proxy without sharing keys.

Agent isolation and first-run setup

On Linux, Agent J runs the phone's Agent inside bubblewrap; socat is what Claude Code's own sandbox needs. The one-line installer installs both with apt-get; when it cannot, it prints the command to run yourself: sudo apt-get install -y bubblewrap socat (dnf on Fedora, pacman on Arch). Without them, every start of the Agent needs the owner's passkey on the phone for "Run this launch without isolation".

When Claude Code is selected, the script also answers its one-time setup ahead (onboarding, theme, bypass-mode confirmation), so a background session never stops at a terminal prompt nobody sees — only what is missing, never what you set. As root it does not write full access (bypassPermissions): Claude Code refuses that for root and exits.

Agent J

At the end, setup asks “现在安装 Agent J 吗？[Y/n]”. Yes opens the existing official installation assistant; it needs your account-page installation code (AJI-…), entered locally with hidden input. --with-agentj selects this step; --no-agentj skips it. These flags choose the step; pairing and account consent still require you. Without a terminal, setup only prints the next step. If no login/key is detected, it explains: “Agent J 能装能配对，但要登录 AI 工具或填 key 后才能真正聊天”. Credential presence is not proof that a token is still valid.

When it asks for Installation code (AJI-…), choose "Copy installation code" on the account page and paste only that AJI- code. Pasting the whole prompt or install command by mistake is fine: the script takes the first AJI code out of it, and full-width characters, spaces and line breaks are handled.

The formal assistant verifies a tool-free model call with installed Claude Code, Codex and OpenCode. It reuses a callable CLI; if several work, you choose. After phone pairing, send Hello. It asks for a formal service URL/model and a secure phone key card only if none can call. AgentsRelay configuration works in the background service without opening a new terminal. Gemini is installed only when explicitly selected and cannot be selected as the main Agent in Agent J. If only Gemini is installed, the assistant asks you to sign in to Codex / Claude Code / OpenCode or use AgentsRelay.

Please download https://agentj.app/install.md with curl, read the whole file with your file-reading tool, then follow it step by step, re-reading each section before you do it. Tell me when you need me to do something.

Installation steps · install.md · Command reference

Keep your computer awake: commands, system settings and a request to send your Agent.

Kimi Code CLI

Kimi Code CLI (kimi, Moonshot AI, MIT) is an AI coding tool Agent J can use directly. The official script needs no Node.js:

curl -fsSL https://code.kimi.com/kimi-code/install.sh | bash

On Windows (PowerShell): irm https://code.kimi.com/kimi-code/install.ps1 | iex (install Git for Windows first). With npm instead: npm install -g @moonshot-ai/kimi-code (in mainland China add --registry=https://registry.npmmirror.com). Check with kimi --version — Agent J needs 2.2.0 or newer.

Sign in in your own terminal: kimi login --region mainland-cn for kimi.com, kimi login --region global for a membership bought on kimi.ai. It prints a link and a one-time code; confirm on any device signed in to Kimi. Then agentj agent switch kimi.

Isolation: Kimi has no sandbox of its own and its hooks let calls through when they fail, so Agent J runs it inside the same isolation as the other tools, keeps it in manual-approval mode and sends every permission request to your phone.

Alibaba Cloud Bailian Token Plan (mainland China)

A monthly plan for coding tools instead of pay-as-you-go API prices. Token Plan, personal, has four tiers — Lite, Essential, Standard and Pro — metered in Credits (11,500 / 25,500 / 45,000 / 180,000 a month; Standard and up run 3–4 agents at once). Your Agent shows the monthly prices with agentj agent route; Alibaba Cloud's subscription page is authoritative. Qwen 3.8 Max and Qwen 3.8 Flash read images, so screenshots and browser work are fine; Agent J picks qwen3.8-max and offers qwen3.8-flash to save Credits.

Read its terms before you buy; the Alibaba Cloud subscription page is authoritative. Your own scheduled tasks and background Agents may use the Token Plan key. The provider may rate-limit long, continuous background work; if that happens, switch those tasks to a pay-as-you-go key. While you are subscribed, what you send to the model and what it writes are used to improve Alibaba Cloud's models, so be careful with private data. The plan is for you only; do not share it. A support bot serves other people, so Agent J never gives it the Token Plan key: add a pay-as-you-go key for the bot (its settings page tells you where).

Subscribe at https://bailian.console.aliyun.com/cn-beijing/subscription/overview and create the plan's own API key in the console. Then ask your Agent to set it up. It runs agentj provider bailian, which adds Bailian to OpenCode's settings and leaves everything you already have there untouched. A key card then opens on your phone: paste the key there and nowhere else. Agent J checks it against Bailian's read-only model list, which costs no Credits. Finally your Agent switches OpenCode to bailian-token-plan-personal/qwen3.8-max.

The Coding Plan (keys starting with sk-sp-) also works (agentj provider bailian --plan coding-plan), but it has no Max model and its key cannot be checked in advance. Your own scheduled tasks may use the Coding Plan key (the provider may rate-limit long, continuous background work; switch those to a pay-as-you-go key if it does); a support bot never gets it and needs a pay-as-you-go key.

Qwen Code

Qwen Code (qwen, Alibaba, Apache-2.0) is Alibaba's own coding agent, and Agent J can drive it directly with the same Bailian key. Install it with npm (Node.js 20 or newer; in mainland China add --registry=https://registry.npmmirror.com):

npm install -g @qwen-code/qwen-code@latest

Check with qwen --version; Agent J needs 0.25 or newer. Its free sign-in ended on 2026-04-15, so the key is its login: the same phone key card as above (BAILIAN_TOKEN_PLAN_API_KEY). Agent J hands the key to Qwen Code in memory only and writes nothing into ~/.qwen. If you already set Qwen Code up yourself, for example with bl config agent --agent qwen-code, your own settings are used. Then agentj agent switch qwen.

Isolation and approvals: Agent J runs Qwen Code inside the same isolation as the other tools. Qwen Code normally starts in a mode where it approves some actions by itself; Agent J switches it back to asking every time, and every request goes to your phone, for that one action only. On the phone, /model lists the plan's models; the ones marked as reading images can look at screenshots.

Volcano Ark Coding Plan (mainland China)

Who it is for: our first recommendation in mainland China if you have no plan yet, and anyone who already has one. One subscription covers ByteDance Doubao, Kimi, Zhipu GLM, DeepSeek and MiniMax models. Buy Lite first: Lite and Pro have the same models and tools, and Pro only gives 5x the quota, so upgrade only if Lite runs short. agentj agent route shows the list prices; the order page is authoritative.

Which models read images (measured on 2026-10-15): doubao-seed-2.1-pro (the default), kimi-k3 (uses quota fast; Ark suggests it for Pro, but Lite can pick it too), kimi-k2.7-code, glm-5-3-flash, minimax-m3 and deepseek-v4-1-flash can look at screenshots; glm-5.3, deepseek-v4-pro and deepseek-v4-flash are text only. Agent J writes this into OpenCode's settings for you, together with higher output limits (GLM-5.3's thinking can use up OpenCode's 32k default). The Coding Plan has no image-generation model.

It is a monthly plan; Ark's console terms are authoritative. Your own scheduled tasks and background Agents may use the Coding Plan key. The provider may rate-limit long, continuous background work; if that happens, switch those tasks to a pay-as-you-go Ark key. A support bot serves other people, so Agent J never gives it the Coding Plan key: add a pay-as-you-go key for the bot (the bot's settings say where).

Sign up at https://www.volcengine.com/activity/codingplan with your phone number, verify your identity and subscribe to the Coding Plan. Then create an API key in the Ark console (https://console.volcengine.com/ark/) and ask your Agent to set it up. For OpenCode it runs agentj provider ark, which adds all the plan's models to OpenCode's settings and leaves everything you already have there untouched. A key card then opens on your phone: paste the key there and nowhere else. Ark has no read-only check, so the first reply is the check. Finally your Agent switches OpenCode to ark-coding-plan/doubao-seed-2.1-pro.

Prefer Claude Code? Ask your Agent to set Ark up for Claude Code instead; it follows the three steps Agent J prints, with the same phone key card. Claude Code's Haiku, Sonnet and Opus defaults and its subagent model all point at the same Ark model, so background work never asks Ark for a Claude model name.

DeepSeek dsh

dsh (DeepSeek Harness, MIT) is DeepSeek's own coding agent, still a developer preview. It is also the engine inside Lares, the agent that comes with an Olares box, so on Olares it pairs naturally with the box's own Router. Install it with npm (Node.js 22.19 or newer; about 500 MB):

npm install -g @deepseek-ai/dsh@latest

Check with dsh --version; Agent J needs 0.2.0-rc.2 or newer. Then choose where its model comes from:

Your own model server (Olares Router, LiteLLM, Ollama, llama.cpp, vLLM or any OpenAI-compatible server): ask your Agent to run agentj agent dsh --dir <folder> --base-url https://<server>/v1 --model <model id>. Agent J saves the address and the model in ~/.dsh/agentj-route.yml; no key is written there. Plain http:// works only for this computer or your home network. If the server needs a key, a key card opens on your phone (DSH_ROUTER_API_KEY): paste it there and nowhere else. Then just send your message again; no restart.
DeepSeek's own API (pay as you go): the same kind of phone key card, for DEEPSEEK_API_KEY.
You already set dsh up yourself (for example on the Models page of dsh web): your own settings are used.

Isolation and approvals: Agent J starts dsh read-only, asking every time, so every file change or command that is not read-only becomes a phone card, for that one action only. dsh's own usage reporting is switched off. dsh cannot look at images yet, and DeepSeek says the preview may still change.

这一页的 Markdown 原文This page as Markdown: zh.md · en.md