npx-vibe

Evidence-first safety checks for npm packages and project dependencies before developers, coding agents, or MCP clients execute them.

npx-vibe resolves packages from the public npm registry, downloads and verifies their tarballs without executing them, inspects install-time code, and prints clear Proceed, Caution, or Block verdicts. Use it for one-off npx commands or scan an existing project's direct dependencies.

The default scan is deterministic, local, and requires no account or API key. AI review is optional, opt-in, and lets you choose an exact model or a maintained fast, balanced, or strong profile. Version 1.5 also ships a zero-dependency MCP server with read-only package, project, and model-catalog tools.

• npm package

• Official MCP Registry entry

• Live website

• GitHub repository

• Security policy

• Changelog

• Release process

Live adoption

The project website displays the current seven-day download total directly from npm's public download API:

• weekly total

• daily seven-day range

The counter is live rather than hard-coded, and the site links to the source data. npm download counts measure package fetches, not unique users. They are momentum signals—not proof of package safety or quality—but they show that the package is being discovered and installed.

Related MCP server: Dependency Checker MCP Server

Quick start

Review a package without executing it:

npx npx-vibe esbuild

Use it as a guarded replacement for npx:

npx npx-vibe cowsay -- hello from npx-vibe

Or install the command globally:

npm install -g npx-vibe
npx-vibe <package>

Scan an existing project's direct dependencies:

npx npx-vibe --project .

Give a coding agent a stable, read-only result:

npx --yes npx-vibe@latest --agent esbuild
npx --yes npx-vibe@latest --agent --project .

Or connect the native MCP server:

npx --yes npx-vibe@latest --mcp

Use npx-vibe with coding agents

--agent turns the scanner into a predictable machine interface. It implies check-only mode, writes only JSON to stdout, disables terminal color and local review-memory writes, and never executes package code. Operational failures are JSON too, so an agent can fail closed instead of interpreting partial terminal output.

Install the portable Agent Skill for Codex, Claude Code, Cursor, VS Code, and other compatible agents:

npx skills add Devrajsinh-Jhala/NPM-Vibe-check --skill npx-vibe -g

The skill tells an agent to preflight unfamiliar packages before npx, npm exec, or dependency installation and then apply the normalized decision:

Example envelope:

{
  "schemaVersion": 2,
  "tool": { "name": "npx-vibe", "version": "2.0.0" },
  "kind": "package-scan",
  "status": "complete",
  "decision": {
    "verdict": "caution",
    "riskScore": 43,
    "action": "review",
    "exitCode": 2,
    "mayContinue": false,
    "safeToExecute": false,
    "requiresApproval": true,
    "requiresHumanReview": true,
    "blocked": false,
    "mustStop": false
  },
  "subject": {
    "type": "package",
    "name": "esbuild",
    "requested": "latest",
    "version": "0.28.1"
  },
  "report": {}
}

The complete deterministic report remains under report. Require schemaVersion === 2, status === "complete", and decision.mayContinue === true before continuing automatically. The outer npx --yes only suppresses npm's download prompt; agent mode rejects npx-vibe's execution flags such as --force and --yes.

AI remains off by default in agent mode. If a user explicitly requests model interpretation, use a provider-specific environment variable and add --ai online --provider <provider>; never place a key in a generated command.

Connect npx-vibe through MCP

Version 1.5 includes a read-only Model Context Protocol server over stdio. It gives MCP-compatible AI applications a native tool surface without teaching them to parse terminal text or construct shell commands.

Add this server configuration to an MCP client that supports local stdio servers:

{
  "mcpServers": {
    "npx-vibe": {
      "command": "npx",
      "args": ["--yes", "npx-vibe@latest", "--mcp"]
    }
  }
}

Global installations can use npx-vibe-mcp as the command with no arguments. Run npx npx-vibe --mcp --help for server-specific help.

The server exposes four schema-backed tools:

Scan results return the same versioned decision contract in both structuredContent and a JSON text block for client compatibility. The tools are annotated read-only and non-destructive. review still requires human approval, while stop, retry, and MCP tool errors fail closed.

AI remains off unless a tool call explicitly selects it. API keys are intentionally excluded from MCP tool arguments: configure provider-specific environment variables on the MCP server process so credentials do not enter prompts or tool history.

npx-vibe@2.0.0 is active in the official MCP Registry as io.github.Devrajsinh-Jhala/npx-vibe. Registry clients can resolve the verified npm package and start its local stdio server with --mcp.

Why developers use it

Running npx some-package can download code and execute a package binary immediately. Packages may also declare lifecycle scripts that run during installation.

npx-vibe inserts a visible checkpoint before execution:

1. Resolve the exact npm package version.

2. Fetch registry, download, maintainer, publisher, and repository context.

3. Download the tarball without running package code.

4. Verify npm integrity metadata and inspect bounded source files.

5. Show matched evidence and a risk verdict.

6. Execute only after the verdict allows it.

For an existing app, --project repeats the same read-only review across direct registry dependencies and aggregates the result. This makes the check useful in local development and pull-request CI, not only before an unfamiliar npx command.

By default, execution uses npm with install scripts ignored. Use --allow-install-scripts only when you intentionally want reviewed root lifecycle scripts to run.

Evidence, not mystery scores

Every deterministic source finding includes the matched line and a bounded excerpt. Registry popularity is displayed separately as context and never overrides suspicious code. Local review memory is keyed by the verified package integrity—not merely its name or version. The example below mirrors a real scan; registry dates and download counts naturally change over time.

$ npx npx-vibe esbuild
! npx-vibe: Caution  risk 43/100
esbuild@0.28.1

Downloads: 241,858,907/week  Package age: 3132d  Version age: 12d
Install hooks: postinstall
Established signals: long registry history, high weekly adoption, linked GitHub repository
Registry popularity and age provide context, but never override code findings.
Review memory: unchanged tarball since 2026-06-25; previous Caution 43/100.
AI review: skipped (Heuristic-only mode; AI was not requested.)

Findings:
- MEDIUM   lifecycle_hook in package.json
  postinstall runs: node install.js
  Evidence: postinstall: node install.js
- MEDIUM   network_and_shell in install.js
  Code combines network access with shell execution.
  Evidence line 147: fetch(url) ... child_process.execSync(...)

Action: review recommended before execution.

A popular package can still receive Caution when it performs sensitive install-time behavior. That is intentional: maturity is useful context, not a security exemption.

When AI is explicitly enabled, the resolved provider and model are visible in the result. This example is adapted from the successful Gemini 3.5 Flash run on June 25, 2026; model wording can vary:

$ npx npx-vibe --ai online --provider gemini --model gemini-3.5-flash esbuild
! npx-vibe: Caution  risk 43/100
esbuild@0.28.1

Install hooks: postinstall
AI review: Gemini gemini-3.5-flash (high confidence)
AI evidence: 0 source-backed findings

AI interpretation: The selected install script appears to resolve a platform-specific binary.
No additional source-backed credential access, obfuscation, or persistence finding was
identified, but the deterministic install-time network and process evidence remains.

Action: review recommended before execution.

What it checks

Built for repeated use

npx-vibe always performs a fresh registry lookup, tarball download, integrity verification, and deterministic scan. Local review memory adds comparison context after those checks:

Review memory: unchanged tarball since 2026-06-25; previous Caution 43/100.

When the integrity changes:

Version comparison: 0.28.1 → current; integrity changed.
2 selected files changed; findings +1/-0; install hooks unchanged.

The history file stores package versions, integrity hashes, selected-file hashes, finding identifiers, verdicts, and model metadata. It does not store package source, API keys, environment values, or local project files. The default location is ~/.npx-vibe/reviews.json.

Verdicts

Registry context — package age, version age, and download volume — is reported at the info severity and never contributes to the score. It describes how new a package is, not whether it does anything.

A verdict is a decision aid, not proof that a package is safe or malicious.

Usage

npx-vibe [options] <package-spec>                      # scan a package
npx-vibe run [options] <package-spec> [-- arguments]   # scan, then execute
npx-vibe approve-scripts [options] [--write]           # settle npm 12 allowScripts
npx-vibe --project <directory|package.json> [options]  # scan a dependency tree
npx-vibe --mcp

Scanning never executes package code. npx-vibe run is the only command that does, and it still refuses a Block verdict unless you pass --force.

Common commands:

# Scan a package; nothing is executed
npx npx-vibe <package>

# Machine-readable output
npx npx-vibe --json <package>

# Versioned, fail-closed JSON for coding agents
npx --yes npx-vibe@latest --agent <package>

# Review, then execute when permitted
npx npx-vibe run <package> -- <arguments>

# Select one executable from a package with multiple binaries
npx npx-vibe run --bin tsc typescript -- --version

# Decide which dependencies may run install scripts under npm 12
npx npx-vibe approve-scripts
npx npx-vibe approve-scripts --write

# Scan production and optional dependencies from this project
npx npx-vibe --project .

# Scan only the direct dependencies named in package.json
npx npx-vibe --project . --direct-only

# Include direct development dependencies
npx npx-vibe --project . --include-dev

# Machine-readable project report
npx npx-vibe --project . --json

# Agent-ready project report
npx --yes npx-vibe@latest --agent --project .

# Start the MCP server over stdio
npx --yes npx-vibe@latest --mcp

# Execute a Caution verdict without prompting
npx npx-vibe run --yes <package>

# Execute a Block verdict intentionally
npx npx-vibe run --force <package>

Useful options:

--json
--agent
--mcp
--project <path>
--include-dev
--direct-only
--max-packages <1-5000>
--no-advisories
--ci
--concurrency <1-8>
--ai-limit <0-100>
--yes, -y
--force
--bin <name>
--ai off|auto|online|ollama
--provider auto|openai|anthropic|gemini|openrouter|groq|together|custom
--models
--model <name>
--api-key <key>
--api-url <url>
--ollama-url <url>
--ollama-model <name>
--registry <url>
--age-days <days>
--downloads <count>
--caution-score <0-100>
--block-score <0-100>
--allow-install-scripts
--no-history
--history-file <path>
--check                     (deprecated; scanning is the default and this is a no-op)
--no-color

Run npx npx-vibe --help for the complete CLI reference.

Project dependency scans

Project mode turns the one-package review into a repeatable dependency preflight:

npx npx-vibe --project .

It reads package.json and, when present, package-lock.json locally. Exact versions from npm lockfiles are preferred over version ranges. By default it scans direct dependencies and optionalDependencies; add --include-dev for devDependencies.

Every registry-resolved package in package-lock.json is reviewed, not only the direct ones, because supply-chain compromises usually arrive transitively. Pass --direct-only for the narrower, faster check:

npx npx-vibe --project . --direct-only

Without a lockfile the scan degrades to direct dependencies and says so, rather than failing.

! npx-vibe project: Caution  highest risk 43/100
my-app@1.0.0

Scanned: 12/12 direct dependencies  Proceed: 11  Caution: 1  Block: 0
Scope: dependencies + optionalDependencies
Resolution: exact versions from package-lock.json when available
AI review: off (heuristic-only)

Packages:
- CAUTION  43/100  esbuild@0.28.1
  young_version, lifecycle_hook, network_and_shell

No dependency or package code was executed during this project scan.
Action: review the flagged dependencies and their evidence individually.

The workflow is deliberately bounded:

• The whole installed tree from package-lock.json is scanned by default. --direct-only narrows it to the dependencies named in package.json.

• --max-packages (default 500) caps how many packages one scan reviews. Anything past the cap is reported as skipped, never silently dropped.

• Workspace, local, alias, URL, and Git specs are reported as skipped rather than uploaded or resolved through another trust path.

• Heuristic-only scans use three concurrent reviews by default (--concurrency 1-8). Weekly download counts for the whole scan are fetched through npm's bulk endpoint in a single request.

• If AI is opted in, reviews are sequential and only heuristic-triggered packages call the model. The default budget is three calls (--ai-limit 0-100).

• package.json and package-lock.json are never sent to an AI provider. Optional online AI receives only bounded files selected from the downloaded registry package.

This is autonomous triage rather than autonomous execution: discover, resolve, verify, inspect, escalate when requested, and aggregate. It never installs dependencies or edits the project.

AI is optional and opt-in

Ambient keys such as OPENAI_API_KEY or GEMINI_API_KEY do not activate AI in the default mode.

Important: Do not paste long-lived API keys into screenshots, issues, chat messages, or shared terminal recordings. Revoke any exposed key immediately.

npx npx-vibe <package>                         # heuristic-only default
npx npx-vibe --provider gemini --api-key <key> <package>  # direct-key shortcut
npx npx-vibe --ai online <package>             # use an explicitly configured online provider
npx npx-vibe --ai auto <package>               # detect configured provider or local Ollama
npx npx-vibe --ai ollama <package>             # local Ollama

Provider-specific keys are read only after --ai online or --ai auto is selected:

OPENAI_API_KEY=... npx npx-vibe --ai online <package>
ANTHROPIC_API_KEY=... npx npx-vibe --ai online <package>
GEMINI_API_KEY=... npx npx-vibe --ai online <package>
OPENROUTER_API_KEY=... npx npx-vibe --ai online <package>
GROQ_API_KEY=... npx npx-vibe --ai online <package>
TOGETHER_API_KEY=... npx npx-vibe --ai online <package>

Provider-specific environment variables are the safest and most reliable option because they avoid provider guessing and keep secrets out of shell history. Recognizable direct-key formats can be routed automatically, but ambiguous formats stop locally and ask for --provider rather than sending a credential to a guessed service.

Google introduced new Gemini authorization keys in June 2026. npx-vibe 1.2.0 recognizes both the newer authorization-key family and traditional Google API keys, and sends Gemini credentials using Google's documented x-goog-api-key header.

PowerShell example:

$env:GEMINI_API_KEY="<new-key>"
npx npx-vibe --ai online --provider gemini esbuild
Remove-Item Env:GEMINI_API_KEY

Direct-key example when you intentionally want to specify the provider:

npx npx-vibe --provider gemini --api-key "<new-key>" esbuild

If automatic routing cannot confidently identify a direct key, npx-vibe exits locally with instructions to add --provider. It does not try the key against OpenAI or any other guessed endpoint.

Custom OpenAI-compatible endpoint:

npx npx-vibe --ai online \
  --provider custom \
  --api-url https://models.example.com/v1/chat/completions \
  --api-key <key> \
  --model <model> \
  <package>

Online AI receives bounded package metadata, deterministic findings, install scripts, and selected files from the downloaded package tarball. It does not receive your project files, shell history, npm tokens, or environment-variable values.

AI findings are checked against the inspected source before they are displayed as evidence. A model finding records its file, line, exact excerpt, and rationale. Unsupported model claims are omitted from the source-backed findings section, and an unsupported AI recommendation cannot independently produce a Block verdict.

Naming the model is required

npx-vibe ships no model catalog. A pinned model id rots: when a provider retires one, the call fails, the review becomes ai_unavailable, the score is floored, and the package reports a false Caution. A scanner whose accuracy decays on a timer is worse than one that asks you to name the model.

npx npx-vibe --ai online --provider anthropic --model <id> <package>
npx npx-vibe --ai online --provider gemini --model <id> <package>

--models lists every supported provider, the environment variable it reads its key from, and where that provider publishes its current model list:

npx npx-vibe --models

Official model lists: OpenAI, Anthropic, Gemini, OpenRouter, Groq, and Together.

Local models need no key and keep their own default:

npx npx-vibe --ai ollama --ollama-model qwen2.5-coder <package>

Automation, agents, and CI

Use JSON plus exit codes in local automation:

npx npx-vibe --json <package> > npx-vibe-report.json
npx npx-vibe --project . --json > npx-vibe-project-report.json

Use the versioned agent envelope when another tool or coding agent owns the decision loop:

npx --yes npx-vibe@latest --agent <package>
npx --yes npx-vibe@latest --agent --project .

Agent mode keeps stdout machine-readable for successful, incomplete, and failed scans. It is deliberately incompatible with --force, npx-vibe's --yes, --allow-install-scripts, and package execution arguments.

npm 12 migration guide

npm 12 turned install scripts off by default, and the hard part is not getting the list of pending packages — it is deciding. npm 12 allowScripts: how to decide what to approve covers the three questions that matter, what the eight most common install-script packages actually do, the .npmrc setting that silently disables the whole allowlist system, and how to enforce the result in CI.

GitHub Action

Listed on the GitHub Marketplace. The action wraps the same CLI, so a workflow needs one step:

name: Dependency preflight

on: [pull_request]

jobs:
  preflight:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
      - uses: actions/setup-node@v6
        with:
          node-version: 24
      - uses: Devrajsinh-Jhala/NPM-Vibe-check@v2

Review install-script permissions instead of the dependency tree, and fail the build when a lockfile change introduces one you have not decided on:

      - uses: Devrajsinh-Jhala/NPM-Vibe-check@v2
        with:
          command: approve-scripts
          fail-on: caution

Outputs are verdict (proceed, caution, block, error) and exit-code. An incomplete scan always fails the step, even with fail-on: never — a scan that could not finish is never read as a pass.

For GitHub Actions, --ci emits a warning for each Caution result, an error for each Block or operational failure, and writes a package table to the job summary:

name: Dependency preflight

on:
  pull_request:
  workflow_dispatch:

jobs:
  npx-vibe:
    runs-on: ubuntu-latest
    permissions:
      contents: read
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with:
          node-version: 20
      - run: npx --yes npx-vibe@2.0.0 --project . --include-dev --ci

Project mode preserves the normal exit contract: 0 Proceed, 2 Caution, 3 Block, and 1 for an incomplete scan caused by an operational error. --ci and --json are intentionally separate so JSON output remains valid.

The repository tests Node.js 20, 22, and 24 across Linux, Windows, and macOS. CI also packs the npm tarball, installs it into a clean temporary consumer project, and exercises the shipped CLI and MCP handshake from the exact artifact users receive.

The tag-driven release workflow is prepared for npm trusted publishing and npm publish --provenance. Configure the trusted publisher once, then push a version tag to publish without storing a long-lived npm token. See RELEASING.md.

Configuration

NPX_VIBE_AI=off
NPX_VIBE_PROVIDER=auto
NPX_VIBE_API_KEY=...
NPX_VIBE_API_URL=https://api.openai.com/v1/chat/completions
NPX_VIBE_MODEL=                # required for online AI review
NPX_VIBE_HISTORY=on
NPX_VIBE_HISTORY_FILE=~/.npx-vibe/reviews.json
NPX_VIBE_OLLAMA_URL=http://127.0.0.1:11434
NPX_VIBE_OLLAMA_MODEL=qwen2.5-coder
NPX_VIBE_AGE_DAYS=14
NPX_VIBE_DOWNLOADS=1000
NPX_VIBE_CAUTION_SCORE=40
NPX_VIBE_BLOCK_SCORE=70
NPX_VIBE_CONCURRENCY=3
NPX_VIBE_AI_LIMIT=3
NPX_VIBE_MAX_PACKAGES=500
NPX_VIBE_ADVISORIES=on         # "off" skips the OSV lookup
GITHUB_TOKEN=                # optional; raises the GitHub metadata rate limit

Supported scope

npx-vibe supports public npm registry package names, scoped packages, dist-tags, exact versions, common semver ranges, and dependency discovery from npm package manifests and lockfiles, including the full transitive tree with --transitive. Version ranges resolve to stable releases unless the range itself names a prerelease, matching npm. It intentionally rejects local paths, arbitrary tarball URLs, Git URLs, and non-registry project dependencies to keep the trust boundary narrow.

Node.js 20 or newer is required. The project is tested on current Windows, macOS, and Linux GitHub-hosted runners.

Security boundary

npx-vibe is a pre-execution risk scanner. It is not a sandbox, antivirus engine, formal audit, or guarantee of safety. A package may hide behavior in unselected files, runtime branches, native code, or remote responses.

Review coverage is bounded and reported rather than implied. Files are selected from the install and executable entry points and the imports reachable from them; each selected file is read up to 64 KiB, and the scan stops selecting once the byte budget is reached. The dashboard names how many files were read only in part or left past the budget, so a large bundled file is never mistaken for a fully reviewed one.

• Treat Proceed as a useful signal, not proof.

• Read the evidence for Caution findings.

• Do not bypass Block unless you understand the behavior.

• Report vulnerabilities privately through SECURITY.md.

• Report noisy findings with the false-positive template.

Contributing

Contributions that improve detection quality, evidence, compatibility, or false-positive handling are welcome. See CONTRIBUTING.md.

License

MIT © Devrajsinh Jhala