
Installed, the system scans...
Before an assistant touches local files, check whether it will install a skill you cannot inspect. WinClaw reviews a skill before it takes effect, scanning for outbound calls, credential reads, and obfuscation, and it rejects high risk. Execution limits require a reason before a sensitive step, keep browser actions visible, and ask again before reading credentials. Missing components are filled in, and model access and parameters are changed in the interface instead of a hand-edited config file. Skill packages can be passed locally between colleagues. It fits people who worry about malicious skills and accidental file changes and still want a graphical desktop agent. The website is mostly a slogan, and a trial asks for a separate form.
WinClaw from Winicssec is a desktop agent. Before a connection, device signatures check identity both ways, and only an instance with a valid key can attach. Before a new skill is installed, the system scans for outbound calls, credential reads, and obfuscated code, and high-risk findings cannot be bypassed. While running, it blocks prompt injection, dangerous commands, and data leaks. Sensitive actions must be explained to the user first, browser actions stay visible, credential reads need a second confirmation, and leftover processes are cleaned up. Security events go to a local audit log. Setup fills in missing components, models are chosen in a graphical screen, the interface ships in Chinese, English, and Japanese, and skill packages can be imported and shared locally.
Difficulty: Intermediate
Identity check before attach
Device signatures check both sides, and only an instance with a valid key can attach, to block an unauthorized process impersonating the agent.
Scan before a skill installs
A new skill is scanned before install. Outbound calls, credential reads, and obfuscation are blocked, and high risk cannot be bypassed.
Message checks while running
Messages in and out are checked for prompt injection, dangerous commands, and data leaks, and a confirmed threat is stopped before it runs.
Confirmation for sensitive steps
Sensitive steps explain their reason first, browser work stays visible, credential reads need a second confirmation, leftover processes are cleaned up, and events go to a local log.
Models in a graphical screen
Setup fills missing components. Model choice and parameters happen in the interface. Skill packages import and share locally. The interface includes Chinese, English, and Japanese.
Scan a third-party skill first
When a person imports a shared skill package, the install scan runs first and stops high-risk packages that call out or read credentials.
Audit local automation
Users who need a trail have file and browser actions explain themselves first, and security events stay in a local log.
The announcement puts the check before install. A new skill's code is scanned, and outbound calls, credential reads, and obfuscation are blocked. High risk is a veto that cannot be skipped. After that, live messages are still checked for prompt injection, dangerous commands, and data leaks. It does not rely on the user reading the whole source after install.
The announcement says a sensitive action must explain its reason to the user first. Browser work stays visible and is not a silent background run. Reading credentials needs a second confirmation. Leftover processes are cleaned up, and security events go to a local audit log. It is not meant to edit files completely unattended. The website homepage does not expand those switches.
Real reviews and feedback from users