pinright
Version-accurate docs and a dependency check for coding agents. Your agent gets library
versions from one of two places: its memory, where about 40% of the pins it wrote in our benchmark
carried a known advisory, or a docs MCP server, which returns documentation and does not check the
version against advisories. pinright gives it the docs for the exact version it installs and a
verdict on every dependency it adds, before the install runs. MCP server, pre-install hook, CLI and
GitHub Action, all local.
Install
Node 20.9 or newer. One line, then restart the agent:
npm i -g pinright && pinright init --user # just you, every project; the hook starts in about 0.2 s
npx -y pinright init # your team: writes project files you commit
init wires the MCP server, the pre-install hook where the client has one, and a starter
.pinright.json. 16 clients; six have a hook: Claude Code, Cursor, VS Code (GitHub Copilot),
OpenAI Codex, Gemini CLI and Windsurf. The Cursor and VS Code buttons above add the MCP server
only. Every client, scopes and pinright doctor.
Without pinright, with pinright
Without. The agent runs npm install [email protected]. It installs. The lockfile now carries
five advisories, two of them high, and nobody is told.
With. The hook answers before the command runs, and the agent reads this:
pinright blocked this dependency change:
- [email protected] (npm) has 5 known advisories (CVE-2021-23337, CVE-2026-4800, CVE-2020-28500, HIGH). Use 4.x release 4.18.1: npm install [email protected]
Change the version (or the name) and try again. check-package and get-docs from the pinright MCP server explain each case.
The agent retries with 4.18.1, on the same release line. When it asks first, check-package
and get-docs give it the same verdict with the full advisory list, the version to pin, and the
docs for that version. Real output of pinright hook claude, 2026-11-01.
Why pinright
Docs MCP servers return documentation for a library. They do not check the version the agent pins,
and they do not see the install. To measure that kind of server we ran pinright's own server with its
security layer off (PINRIGHT_ABLATION=docs-only): docs for the version asked, no verdict. It is not
another product. Claude Code with Sonnet 5.5, Haiku 5.5 and Opus 5.5, 2 runs per cell, on the 1.0.0
code; each cell gives the range over the three models.
| No docs | Docs-only context server | pinright |
|---|
| Direct dependencies pinned with a known advisory | 40% to 42% | 0% to 2% | 0% |
Whole-tree alerts per project (npm audit, pip-audit) | 29.8 to 31.9 | 0.3 to 3.5 | 0.3 |
| Docs tasks that used the current API | 9 to 14 of 16 | 15 or 16 of 16 | 16 of 16 |
| Tokens per agent task | 15k to 21k | 56k to 94k | 42k to 74k |
| Tokens per docs task | 17k to 23k | 94k to 135k | 63k to 95k |
| Agent turns per agent task | 4.0 to 4.3 | 9.3 to 11.1 | 5.9 to 7.5 |
| MCP tool text read per agent task | none | 25k to 45k characters | 5k to 17k characters |
| Verdict on the version the agent pins | no | no | advisories, malware, lookalikes, versions that do not exist |
| Check when the install runs | no | no | pre-install hook in 6 clients |
| Whole-tree lockfile audit | no | no | check-manifest, SARIF, GitHub Action |
Current docs alone removed most alerts; the verdict removed the rest. Per model, pinright used 21% to
26% fewer tokens than the docs-only server on the 12 agent tasks and 26% to 33% fewer on the 8 docs
tasks. Every paired interval stays above zero; the lowest bound is 2%. Tokens count everything the
model processed, cache reads included. No tools at all costs less than either.
Why fewer tokens. With docs only, the agent looked versions up through the docs: it called
get-docs once per package, 5.2 to 5.7 times per agent task, before writing a file. 87% of those calls
named no version, so they returned the docs of the latest release, and 75% asked for "version" or
"latest" in the query. With pinright, every agent-task run opened with check-packages for the whole
list, about 600 characters of verdicts and versions to pin, then called get-docs 0.7 to 1.7 times per
task, with a version in 86 of 87 calls.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
