OpenClaw Cursor Agent

Manage long-running Cursor CLI coding tasks through OpenClaw tools backed by tmux sessions. Use when the user asks to start a persistent coding job, inspect...

Security Scan

The skill's code and scripts generally match its stated purpose (managing tmux-backed Cursor tasks), but there are important mismatches and privilege/exposure risks you should be aware of before installing.

Purpose & Capability

The plugin, JS entrypoint, and the included shell scripts are coherent with the described purpose: they spawn tmux sessions, run Cursor CLI, track status files, send commands, and kill sessions. However, the package metadata claims "Required binaries: none" whereas the runtime scripts and docs clearly require bash, tmux, python3, and the Cursor CLI (agent/cursor-agent). That metadata omission is an incoherence and could mislead users about runtime requirements.

Instruction Scope

SKILL.md and the included scripts instruct the agent to run local shell scripts that create task/runner/log/status files and execute Cursor CLI in project directories. Those scripts will run commands in user-specified project paths and can modify the project (git operations are part of recommended tasks). The Node extension spawns child processes with env: process.env (it forwards the entire environment) so any environment variables available to the agent are also visible to the scripts/child processes. The docs also include a restore script and backup download steps that would copy many files into ~/.openclaw if followed — a potentially destructive operation if executed without review.

Install Mechanism

There is no external install spec (no arbitrary download/extract). The skill is instruction/code-only and includes local scripts and a Node entrypoint. No third-party URL downloads or installation of remote archives are present in the plugin files themselves (the docs show curl usage for installing Node/Cursor or downloading backups, which are standard but require user discretion).

Credentials

The skill declares no required environment variables or primary credentials, which superficially looks safe. In practice the Node code launches shell scripts with env: process.env (inheriting the agent process environment) and the scripts run external tools and may operate in arbitrary project directories. That means secrets present in the agent environment (AWS keys, tokens, etc.) could be visible to child processes or to any code the Cursor agent executes. Also the skill expects configuration (toolkitRoot) but that is not declared in the top-level 'required' metadata, increasing the chance of misconfiguration.

Persistence & Privilege

always is false. The skill writes state, task and log files under the configured toolkitRoot/status, tasks, and logs directories (normal for this functionality). It does not request to modify other skills or global agent settings. Nevertheless, scripts and the Cursor agent will be capable of making changes in target project directories (including git commits/pushes) — this is expected for a background coding agent but is a high-impact capability that users must accept explicitly.

What to consider before installing

What to check before installing: - Understand the power: this skill runs shell scripts and the Cursor CLI in your filesystem and can modify project directories (it may run git commands, create files, push changes if instructed). Only install if you trust the code and intended tasks. - Metadata mismatch: the package metadata claims no required binaries, but runtime scripts require bash, tmux, python3, and the Cursor CLI (agent). Expect to install/configure these first. - Environment leakage: the Node extension launches scripts with the full process environment (process.env). Remove secrets from the agent environment or run the plugin under a separate, limited account/environment if you want to reduce risk. - Inspect toolkitRoot: the plugin requires a toolkitRoot pointing at the included cursor-agent-system. Make sure it points to the copy you reviewed. Don’t point toolkitRoot to an existing system path you don’t control. - Review restore/backup steps: documentation contains commands that copy backups into ~/.openclaw and download archives — do not run those blindly; review restore.sh and any backup contents before executing. - Run doctor and tests in an isolated environment: before trusting it with real projects, configure toolkitRoot in a sandbox directory, run openclaw cursor-agent-doctor, and perform the minimal smoke tests described in the docs. - If you need stronger guarantees: run the plugin inside a VM/isolated container or under a user with limited privileges, and avoid exposing sensitive env vars to the OpenClaw process.

让 OpenClaw 把长时间编码任务交给 Cursor CLI 在 tmux 里持久运行。

架构

用户（飞书/聊天） → OpenClaw Agent → 插件 → WSL → tmux → Cursor CLI Agent

可用工具

工具	说明
cursor_agent_spawn_task	启动后台 Cursor 编码任务
cursor_agent_list_tasks	列出所有任务
cursor_agent_check_status	查询任务状态和进度
cursor_agent_send_command	向任务发送补充指令
cursor_agent_kill_session	结束任务
cursor_agent_doctor	诊断环境和依赖

工作流程

1. 环境不确定时先 cursor_agent_doctor 检查
2. cursor_agent_spawn_task 启动后台任务
3. cursor_agent_check_status 或 cursor_agent_list_tasks 监控进度
4. cursor_agent_send_command 发送补充指令（/pause、/resume、/status、或自然语言）
5. cursor_agent_kill_session 结束任务

自然语言映射

用户说	调用工具
帮我用 Cursor 在后台做一个任务：…	cursor_agent_spawn_task
看看刚才那个任务做到哪了	cursor_agent_check_status
告诉刚才那个任务：改成 RS256	cursor_agent_send_command
把刚才那个 Cursor 任务停掉	cursor_agent_kill_session
列出所有后台任务	cursor_agent_list_tasks
检查 Cursor 环境是否正常	cursor_agent_doctor

输出规范

• spawn 后告知会话名和任务 ID
• 查状态时报告当前步骤、进度百分比、tmux 是否存活
• 环境缺依赖时明确指出缺少什么（tmux / python3 / agent / WSL）

参考文件

• 插件入口 — 工具注册和执行逻辑
• 工具脚本 — spawn / check-status / send / kill / attach
• 使用指南 — 用户自然语言操作说明
• 配置示例 — openclaw.json 配置参考