The security scanner your coding agent runs before every commit.

Your agent scans, reads the JSON recommendation field, applies each fix, and re-scans until it comes back clean. No dashboard to open. Your agent gets an exit code and JSON, and a report saves to .shipguard/.

Same code → same score → same exit code. Every time, on every machine. An LLM gives three different answers to the same question — ShipGuard gives one, and your agent can trust it enough to branch on it without asking a human.

Free scans need no account and run on your machine. Cloud scans (any signed-in plan) connect once to download rule bundles via an authenticated edge gateway, then scan locally. Either way, no code, file paths, or file contents are ever transmitted. Rules come down. Code never goes up. Published builds (0.10.0 and later) send three small pseudonymous usage pings (after each scan, on first run, and after shipguard login) with no code, file contents, paths, repo names or secrets. Opt out of all three with SHIPGUARD_TELEMETRY=0 or DO_NOT_TRACK=1 or shipguard telemetry disable.

ShipGuard's rules are tuned to Next.js App Router routes, Supabase RLS migrations, Stripe webhook handlers, and Firebase security rules — regex and string matching that knows route.ts in app/api/ is a public endpoint, not just a file with a string in it.

No config files. No account. No setup wizard. One npx command and ShipGuard scans your project — locally, privately.

Every coding agent runs a loop: reason → act → observe → repeat. ShipGuard is the observe step — the deterministic checkpoint between "the agent wrote code" and "the code ships." It returns structured findings the agent can reason about and act on in the next iteration, closing the loop without a human in the middle.

A CLAUDE.md instruction is a hope. A hook runs every time. Wire ShipGuard as a PreToolUse hook or a pre-push git hook and it fires deterministically at the boundary between "agent wants to push" and "code leaves your machine." Exit 1 blocks (with --strict). The agent doesn't get a vote.

Goal-based loops need an objective stop condition — not "does this look safe?" but a binary yes/no the agent can check deterministically. Passed with --strict, ShipGuard's exit code is that condition: exit 0 means no blocking findings remain. The agent's goal becomes: "get to exit 0." It can iterate autonomously — scan, fix, re-scan — with a clear, machine-verifiable definition of done.