Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Stroq: audit what your coding agent touched

Written November 6, 2026

Stroq audits coding-agent transcripts for secrets and provenance, with pre-action hooks. What it checks, and what its attack suite shows.

The problem: nobody checks what the agent read or sent

A coding agent reads files, fetches pages, and runs commands. Its transcript can carry secrets into logs, and an action can come straight out of text the agent read — an issue, a fetched page, a file in the repo — without anyone tracing it back. Stroq is a local security tool for that gap: it reads your Claude Code, Codex, and Cursor sessions, shows how many commands came from text the agent read, and which keys left the machine. It is open source, Apache-2.0, and sends nothing anywhere — no account, no telemetry.

What it checks

Stroq works two ways. After the fact, stroq replay reads the agent's own transcript — including sessions from before Stroq was installed — and shows which content the agent read and which later actions came out of it. stroq sent compares sessions with the credentials on your machine and says where each key went: sent to another company, maybe sent, seen by the AI, used with its own service, or blocked before it ran. It prints key names, never values.

Before the fact, stroq init wires hooks into the agent that judge every tool call before it runs. The hooks deny curl | sh and outbound calls carrying a known secret, and ask you first when a command repeats text the agent read, naming that text. Hooks exist for Claude Code, Cursor, Codex, Copilot CLI, Windsurf, and Antigravity, with a plugin for OpenClaw and a stdio MCP proxy. A local console on 127.0.0.1 shows recorded decisions as they happen, with the chain behind each one: what the agent read, what it did, the decision, and the rule.

Running it is one command: npx @stroq/cli prints a card of your last 30 days of sessions. It needs Node.js 22 or newer.

What the attack suite shows

Stroq ships a built-in attack suite: stroq attack runs 36 known attacks against your policy in a throwaway home. Running it on version 0.30.0 gave the same result the project claims for that version: 15 attacks blocked, 21 put to the user as a question, and 0 passed through. The 36 cover real incidents — a curl | sh from a fetched page, a malicious skill installer, a poisoned issue title piped to a shell — and synthetic cells for classes with no public incident yet, like an instruction hidden in a filename or in invisible characters in an agent instruction file.

That result is a check of the policy against its own test set, not proof the tool stops real attacks on your machine. The attacks run in a throwaway home, not against your sessions, and the suite is written by the same people who wrote the rules it tests. What it does show is that the hooks fire: nothing in the suite slipped past silently.

Where it stops

What to do with it

  1. Run the card first. npx @stroq/cli reads your last 30 days of sessions and prints what it finds. If it shows keys leaving the machine, rotate those keys before doing anything else.
  2. Run the attack suite. npx @stroq/cli attack tells you whether the default policy behaves as claimed on your setup. It runs in a throwaway home and changes nothing.
  3. Wire the hooks in only after the card. stroq init guards the next session. Read the per-agent limits in the docs first — a hook that cannot see what it needs to see is worse than no hook, because it looks like protection.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Quarry: behaviour-chain detection · Audit git config before opening a repo · Check CLI config relay tampering