Open Cowork:带路径沙箱与轨迹视图的开源桌面办公智能体
开源 Electron 应用 Open Cowork 把文件工作区、智能体对话、执行轨迹、文档技能、MCP 与命令沙箱放在同一界面,强调先在隔离文件夹里验证权限边界。
把桌面智能体直接对准整个用户主目录,很难验证它会不会越权改文件。开源项目 Open Cowork 选择另一条路:用 Electron 桌面应用提供受限工作区,把智能体聊天、文件树、执行轨迹(trace)、文档类技能、MCP 连接与命令沙箱放在同一界面,方便观察「从文件到产物」的每一步。
项目支持 Windows 与 macOS,可对接多种模型供应商,并内置面向文档、演示文稿、电子表格与 PDF 的技能。作者强调,真正有用的不是供应商清单,而是可见工作区加上可检查的轨迹:团队若想试用办公类智能体任务,又不愿只盯着终端滚动日志,这种界面更合适。
帖子给出的试用建议很具体:先建只含一份 CSV 与一份无害 PDF 的新文件夹,让智能体在该目录内摘要 PDF、从 CSV 生成小表;再检查轨迹、输出文件与权限提示,并故意尝试写入工作区外路径,确认拒绝是否可观察。作者还说明跨平台路径守卫、Windows 上的 WSL2 隔离、macOS 上的 Lima 虚拟机等模式并不等价,Lima 不可用时可能回退为「本机执行 + 路径限制」,因此要先确认当前会话实际跑在哪种边界下。发行版 3.3.1 增加了本地智能体权限提示、工具调用循环守卫,并修复 Lima/WSL 沙箱中的路径穿越绕过;安全修复是升级与复测策略的理由,不能当作整体无风险的证明。安装指引里若出现绕过 macOS 隔离属性的步骤,也不应随手照抄。
对只要窄场景编码助手的人,Open Cowork 可能偏重;当目标是「可检查的桌面文档与文件自动化」时,适合从一文件夹、一连接器、一输出、一拒绝操作开始,再缓慢扩大范围。
综合来看,读者可先核验官方渠道与发布时间,再结合自身业务边界与合规要求决定是否跟进试点;本文仅整理公开信息,不构成采购、投资或法律建议,关键数字与能力边界请以一手文档为准。
Open Cowork: Bounded Desktop Agent with Trace and Sandbox
The open-source app combines chat, trace, document skills and a sandbox, with boundaries worth testing.
Giving a desktop agent your whole home directory is a bad way to test whether it can make a spreadsheet. Open Cowork offers a more bounded place to try those tasks: a desktop app with a file workspace, agent chat, trace view, document skills, MCP connections and a command sandbox.
The open-source Electron project runs on Windows and macOS. It supports several model providers and includes skills for documents, presentations, spreadsheets and PDFs. The interesting part is not the long provider list. It is the combination of a visible workspace and a trace of what the agent did while moving from files to an output artifact. For a team that wants to test office-style agent work without living in terminal scrollback, that surface is useful.
Try it with dummy data first. Create a fresh folder with one CSV and one harmless PDF. Ask the agent to summarize the PDF and create a small table from the CSV inside that folder. Inspect the trace, the resulting files and every permission prompt. Then test an attempted write outside the allowed workspace. A sandbox claim is only valuable if the denied operation is observable. Keep real client documents, keys and the rest of your home directory out of the first run.
Open Cowork describes path guards across platforms, WSL2 isolation on Windows, and a Lima VM on macOS. When that VM is unavailable, it can fall back to native execution with path restrictions. Those are not identical boundaries. Check which mode your machine actually used before treating the session as isolated. Release 3.3.1 added local-agent permission prompts, a tool-call loop guard and fixes for a path-traversal bypass in the Lima/WSL sandbox. Security fixes are a reason to use a current build and test the policy, not proof that the whole application is risk-free.
One honest note: the install guidance includes a macOS option that bypasses quarantine. Do not copy that step casually. Inspect the release and use a deliberate install path. The app also combines a lot of surfaces, from MCP to remote control. If all you need is a narrow coding agent, this may be more complexity than value.
Open Cowork earns its place when a bounded desktop workflow makes document and file automation easier to inspect. One folder, one connector at most, one output, one denied operation. If that holds up, expand scope slowly. The trace should make the agent's behavior clearer, not merely prettier.