Ship Safe, an open source security scanner for coding agents

Ship Safe is an open-source AI security scanner that hit Hacker News front page as a local tool for finding risky code, AI-agent vulnerabilities, and supply-chain issues before they ship. Teams start it with a single command, npx ship-safe, and get a CLI plus an interactive REPL that can scan, propose fixes, and answer questions in one session. There is no signup and no API key required for scanning.

The product runs in the local repository and inspects application code, AI agents, MCP configs, prompts, dependencies, CI/CD, secrets, and cloud-adjacent configuration. Core checks work offline. Optional AI-backed red-team modes use a configured provider when available; --no-ai forces a fully local scan. Provider-backed classification, deep analysis, and GPT-Red send only bounded context to the selected provider after best-effort credential masking, with exact boundaries and context limits documented under Security and Data Flow.

For engineers building agentic workflows, that scope matters because risk is no longer limited to classic app code and dependency graphs. MCP servers, system prompts, and CI glue are part of the attack surface, and a scanner that treats them as first-class inputs matches how teams actually ship agents. Local-first execution also fits locked-down repos where sending full trees to a SaaS scanner is blocked or undesirable.

The positioning sits against both traditional SAST/SCA tools and newer AI-app security products: Ship Safe markets itself for modern software teams that mix application code with agents and MCP, not only for legacy service code. Open source, free scanning without signup, offline core checks, and optional provider-backed red teaming give it a different cost and trust profile than hosted scanners that require accounts or always-on cloud analysis. Pricing and a public benchmark are listed alongside docs and a contribute path, which signals product maturity beyond a one-off CLI demo.

Practical next steps are concrete: run npx ship-safe in a real agent monorepo, compare results with --no-ai versus provider-backed modes, and read Security and Data Flow before enabling classification or GPT-Red so credential masking and context limits are understood. Watch how the benchmark scores stack up against existing scanners on agent and MCP issues, and whether safe-fix suggestions stay accurate enough to apply without heavy manual review.