FAQ
Questions developers ask first
Does it upload my code?
No. pinright sends package names and versions to public services (deps.dev, OSV, the registries, jsDelivr, GitHub), never your source code, file contents or environment. A private package name you check does reach those services. No telemetry, no account. What goes where
Which ecosystems does it cover?
Verdicts and audits: npm, PyPI, Cargo, Go, RubyGems, NuGet, Maven and Gradle, Composer and GitHub Actions. The hook reads install commands for most of them and manifest writes for all but RubyGems and NuGet. Version-accurate docs cover npm and PyPI packages and GitHub repositories.
What happens if a registry or advisory database is down?
pinright says ok only when deps.dev and OSV both answered. Otherwise the verdict is unknown and names the host. The starter .pinright.json that init writes makes the hook ask you; without a policy file it allows the change and tells the agent it was not verified. A check that runs past 45 seconds asks a human. Answers are cached for 24 hours, docs for 7 days.
How is this different from Dependabot or npm audit?
Those tools report on dependencies already in your lockfile or repository. pinright acts earlier, while the agent picks the version: it names a clean release on the line you use, gives the agent docs for that version, and the hook checks the install before it runs. It also checks lookalike names, malware records and packages that do not exist. For CI, pinright audit and the Action audit the whole tree; on 394 benchmark lockfiles they flagged the same packages as npm audit and pip-audit. Use them together.
Does it slow my agent down?
A little. The hook took 121 ms at the median on a command that installs nothing, 358 ms on one install with a warm cache and 2.4 s cold. In the benchmark, runs with the MCP server took about 2 to 4 more turns and 1.5 to 2 times the API-equivalent cost of a run without tools. A global install starts faster than npx; see the Latency section of the README.
Can the hook be bypassed?
Yes. It reads commands; it is not a sandbox. Code that calls pip from inside a program, curl … | sh, heavy obfuscation and tools the hook is not wired to get past it. What it cannot read asks by default. Keep a proxy registry and run agents in a sandbox. Threat model in brief
Does "ok" mean a package is clean?
It means no advisory or malware record is published for that version, and the name does not look like a typo of a popular one. pinright knows published advisories only. For brand-new attacks it adds supply-chain signals, and the hook blocks two patterns: a package younger than 14 days that runs install scripts, and a release that adds install scripts and comes from a publisher the previous releases never had.
The hook blocked something it should not have. What now?
Add the name to "allow" in .pinright.json to let it through, and please report the false positive with the exact command and the output of pinright doctor --json. It is the report we value most.
What does it cost?
Nothing. No account, no key, no paid tier. The data comes from public services. The only cost is the extra agent turns described above.
What is the license?
MIT, by Hiper AI. Contributions are welcome; start with CONTRIBUTING.md. Questions: opensource@hiperai.com.