AI AgentAI Agent
一句话描述运维目标，Agent 现场规划、逐条执行、看输出再决定下一步——每一步都先经你审批。
它怎么工作
- 下达任务——连接主机后打开 AI 侧栏，输入目标，例如「装好 nginx 并反代到 :3000」。
- 确认计划——Agent 先给出任务理解与步骤清单（支持子步骤嵌套），进度随执行实时更新。
- 逐步审批执行——每条命令附一句执行理由；结果回传模型决定下一步，直到任务完成。
三级审批门
- 只读（
ls/cat/systemctl status…）——自动放行，折叠显示。 - 变更——点击批准后执行。
- 破坏性（
rm -rf/dd/ 重定向系统路径…）——醒目样式 + 二次确认。
命令级别由本地分类器判定，不信任模型自报；你拒绝时可附理由回传，模型据此调整方案而不是傻循环。
改文件先看 diff
写文件走 SFTP（而非 shell 拼接），先展示 diff 预览、批准后才写入。所有命令与输出同步镜像进终端，全程可见。
模型与隐私
- 自带密钥（BYOK）：支持 Anthropic 与 OpenAI 兼容 API，可配置第三方网关地址。
- API 密钥存系统钥匙串；命令输出回传模型前先经本地脱敏（IP、密钥等敏感片段替换为
[REDACTED]）。 - 会话历史本地持久化，可经 iCloud 多端同步；绝不在你不在场时后台执行。
Agent 命令只发当前会话，绝不进入命令广播路径；单命令超时、失败重试与迭代上限都有护栏。
Describe an ops goal in one sentence. The Agent plans on the spot, runs commands one by one and reads the output before deciding the next step — every step waits for your approval.
How it works
- State the task — connect to a host, open the AI sidebar and type the goal, e.g. “install nginx and reverse-proxy to :3000”.
- Review the plan — the Agent first shows its understanding and a step checklist (nested sub-steps supported); progress updates live as it executes.
- Approve step by step — each command comes with a one-line rationale; results go back to the model to decide the next step, until the task is done.
Three-tier approval gate
- Read-only (
ls/cat/systemctl status…) — auto-passes, shown collapsed. - Changes — run after a tap to approve.
- Destructive (
rm -rf/dd/ redirects into system paths…) — prominent styling + double confirmation.
The tier is decided by a local classifier — the model's self-description is never trusted. When you reject, your reason goes back to the model so it adjusts course instead of looping.
File edits show a diff first
File writes go over SFTP (not shell heredocs): you see a diff preview and approve before anything is written. Every command and its output mirror into the terminal, visible the whole time.
Models & privacy
- Bring your own key (BYOK): Anthropic and OpenAI-compatible APIs, with custom gateway URLs.
- API keys live in the system Keychain; command output is redacted locally (IPs, keys and other sensitive fragments become
[REDACTED]) before reaching the model. - Conversation history persists locally and can sync via iCloud; the Agent never runs in the background without you present.
Agent commands go only to the current session — never through command broadcast. Per-command timeouts, retry limits and an iteration cap are always in place.