Image: GitHub
Why it mattersA team running coding agents on real code can audit what each agent did and why. The hooks trace every risky command back to the exact text that caused it.
A coding agent that writes code also writes commands. Some of those commands send secrets to a service, or read files the agent should never touch. Stroq is a local tool that catches those commands.
The project was posted to Hacker News on 5 November 2026 as Show HN. It had 72 points on 6 November 2026, per its own Hacker News page. The repository has 105 stars, 7 forks and 6 open issues on 6 November 2026, per the GitHub API. The licence is Apache 2.0.
What it does
Stroq reads the session logs that coding agents leave on the machine. It shows which commands came from text the agent read. It names which API keys left the machine, and where they went. The project supports Claude Code, Cursor, Codex CLI, Copilot CLI, Windsurf, Antigravity and OpenClaw, the README says.
After reading past work, Stroq installs hooks. The hooks judge each new tool call before the agent runs it. A dangerous call can be denied, flagged for a user to confirm, or allowed. The rules run only on the local machine. The project has no account, no telemetry and no cloud link.
The 36-attack test in the repository
A command called stroq attack runs 36 known attacks against the local policy in a throwaway folder. The README states the result on version 0.30.0: "15 are denied, 21 are asked about, 0 pass through." The attacks come from public reports and synthetic cases.
The number is Stroq's own on its own test suite. An attack list the project picks is one place a user has to trust the maintainer. Reading the test source is cheap, because the project is open source and the attacks sit in the repository.
What a team should check
Running a coding agent unattended is cheap. The cost shows up when the agent leaks a password or deletes work. It also shows up when the agent calls an outside service with a customer's data. A local tool catches the agent on the developer's machine. It stops the leak before it reaches a real server.
Three checks before a team adopts this. First, read the attack list in the repository and name the attacks that match your own risk. A health company has different leaks than a startup. Second, measure the slowdown on a working day. Hooks that judge every tool call add a cost. Third, test what Stroq does when the agent asks a question the hook does not know how to answer. A safe default is to pause, not to allow.
The tool is early. 105 stars means some real users on 6 November 2026. A security tool that watches an unattended agent is only as strong as its rules. The hooks are the real test.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
