Safe npm install: evidence before dependencies

Coding agents can change dependencies faster than a person can inspect a release. Safe npm install puts a consistent decision process before npm, pnpm, yarn, and bun changes and package-safety questions. It uses the free Dep Bouncer MCP server to inspect release history, publisher and provenance changes, install behavior, dependencies, code findings, and advisory evidence.

The agent playbook complements the free server and its bundled prompt: it reconciles incomplete coverage, separates range audits from locked versions, checks peers, explains evidence plainly, and pins the exact checked version. ALLOW permits an authorized install; REVIEW asks the user; BLOCK prevents that release and offers a checked alternative when available. Script decisions remain specific to the package and version.

For Claude Code, Cursor, and generic Streamable-HTTP MCP agents, the package includes setup instructions, a finding reference, and worked examples. It is readable Markdown with no hidden execution, telemetry, credential handling, or bundled enforcement hook.

What you get

• Separate workflows for adds, upgrades, audits, scripts, version selection, and safety questions.
• BLOCK stops and incomplete or uncertain checks require review.
• Minimal manifests are built locally; skipped entries and unaudited peers are reconciled.
• Checked versions are pinned exactly and unnecessary lifecycle scripts stay disabled.
• Signal and code-pattern evidence is translated into practical explanations.

How it works

1. Connect a skill-capable agent to the free hosted Dep Bouncer MCP server or a configured self-hosted server.
2. Choose the workflow for the requested add, upgrade, audit, script review, version selection, or safety question.
3. Send only the tool arguments required for that check; for a manifest audit, first build the allowed minimal manifest locally.
4. Read the structured result, findings, completeness, and coverage notes, then apply ALLOW, REVIEW, or BLOCK rules.
5. For an authorized install, pin the exact checked version, preserve the project’s package manager, and report actual changes and warnings.

Privacy and network calls

MCP calls go to the configured endpoint, by default the author's free hosted service. The server receives package names and requested versions/ranges, or the specific upgrade, script-review, incident, or minimal-manifest arguments for the selected tool, plus ordinary MCP/HTTP metadata. A minimal manifest may disclose the project's name/version and dependency names, values, and section labels; scripts and other project fields are excluded. Private or non-registry dependencies are omitted and treated as unresolved REVIEW. Live checks can query npm registry metadata/tarballs, npm download counts, and OSV advisories. Replays use bundled reconstruction fixtures without live registry/OSV lookups. Installs make normal package-manager requests; approved scripts can make additional calls. The playbook adds no telemetry or uploads of source files, lockfiles, conversation text, environment, or credentials. Checks are npm heuristics, not a safety guarantee.