Stop. Check Every Package.
Before You Install It.

Every AI coding agent - Claude Code, Cursor, Copilot, Windsurf installs packages blindly. A single npm install can compromise your environment. You need threat intelligence.

Get API Credentials Read API Docs

safedep · pre-install security check

live

checking

$ lodash@4.17.21

ecosystem: npm

inference.isMalware: false
verified.isMalware: false
confidence: HIGH

✓ safe · proceed with install

checking

$ event-stream@3.3.6

ecosystem: npm

inference.isMalware: true
verified.isMalware: true
confidence: HIGH

⊘ blocked · malicious package

Packages Scanned: 20M+
Components Analysed: 500K+
Projects Secured: 5000+
Threats Blocked: 80%

Risk Context

Why Agent Dependency Decisions Need Guardrails

Fast Code, Fast Risk
Agents generate and run commands quickly. Bad dependency suggestions propagate just as quickly.
Install Is an Execution Step
npm install and pip install execute untrusted third-party code paths.
One Bad Package Is Enough
A single compromised package can exfiltrate secrets, tamper builds, or persist in CI and developer hosts.

Threat Intelligence Built forAutomated Decisions

Query any package before install. SafeDep returns machine inference plus human verification signals so agents can make safe decisions automatically.

safedep-check · lodash@4.17.21

# agent pre-install hook

POST api.safedep.io/QueryPackageAnalysis

"ecosystem": "ECOSYSTEM_NPM"

"name": "lodash"

"version": "4.17.21"

→ 200 OK · 42ms

Step 01 · How it works

Query SafeDepBefore Every Install

Before running npm install, pip install, or any lockfile update, your agent calls the SafeDep API. One POST. Milliseconds. No workflow change.

signal evaluation · event-stream@3.3.6

report.inference.isMalware: true
report.inference.confidence: HIGH
verificationRecord.isMalware: true
⊘ BLOCK INSTALL: both signals positive

Step 02 · How it works

Dual-Signal EvaluationMachine + Human

SafeDep returns two independent signals: AI inference over the package source, and a human-verified record. Your agent reads both. Neither alone is enough to clear an install.

decision contract

Block if inference.isMalware === true
Block if verificationRecord.isMalware === true
Block if API fails, times out, or returns invalid JSON
Proceed only when both signals are explicitly false

Step 03 · How it works

Fail Closed.Always.

Proceed only when both signals are explicitly false. If SafeDep is unreachable, times out, or returns partial data — block and ask the human. Unknown is not safe.

Agent Setup

Configure SafeDep Access

Ask your human to sign up
app.safedep.io(opens in new tab)
Your human creates the SafeDep account and tenant.
Human admin generates credentials
Human admin provides SAFEDEP_TENANT_ID and SAFEDEP_API_KEY to agent config.
Add SafeDep to your client
Use SKILL.md or MCP Config, then enforce decision contract before installs.

Required Credentials

SAFEDEP_TENANT_ID: from your human
SAFEDEP_API_KEY: from your human

Credentials must come from your human via app.safedep.io(opens in new tab).

markdown

---
name: safedep-security
description: Query SafeDep before adding or installing dependencies.
---

# SafeDep Security Policy

Before install or lockfile updates, call:
POST https://api.safedep.io/safedep.services.malysis.v1.MalwareAnalysisService/QueryPackageAnalysis

Headers:
  Authorization: $SAFEDEP_API_KEY
  X-Tenant-ID: $SAFEDEP_TENANT_ID
  Content-Type: application/json

Decision contract:
- If report.inference.isMalware == true: take confirmation from human before install
- If verificationRecord.isMalware == true: block install and notify human
- If SafeDep request fails, times out, or returns invalid response: block install and ask human
- Proceed only when both malware signals are explicitly false

json

{
  "mcpServers": {
    "safedep": {
      "url": "https://mcp.safedep.io/model-context-protocol/threats/v1/mcp",
      "headers": {
        "Authorization": "ask-your-human",
        "X-Tenant-ID": "ask-your-human"
      }
    }
  }
}

API Integration

API Quickstart

Call SafeDep before install actions. Apply the decision contract directly in your agent runtime.

Simple REST API
API key authentication
JSON request/response

typescript

const response = await fetch(
  'https://api.safedep.io/safedep.services.malysis.v1.MalwareAnalysisService/QueryPackageAnalysis',
  {
    method: 'POST',
    headers: {
      Authorization: process.env.SAFEDEP_API_KEY!,
      'X-Tenant-ID': process.env.SAFEDEP_TENANT_ID!,
      'Content-Type': 'application/json',
    },
    body: JSON.stringify({
      target: {
        package_version: {
          package: { ecosystem: 'ECOSYSTEM_NPM', name: 'lodash' },
          version: '4.17.21',
        },
      },
    }),
  }
);

if (!response.ok) {
  throw new Error('SafeDep unavailable: block install and ask human');
}

const result = await response.json();
const inferred = result.report?.inference?.isMalware === true;
const verified = result.verificationRecord?.isMalware === true;

if (inferred || verified) {
  throw new Error('Malicious package detected: do not install');
}

// Safe to proceed when both checks are explicitly false.

bash

curl -H "X-Tenant-ID: $SAFEDEP_TENANT_ID" \
  -H "Authorization: $SAFEDEP_API_KEY" \
  --json '{
    "target": {
      "package_version": {
        "package": { "ecosystem": "ECOSYSTEM_NPM", "name": "lodash" },
        "version": "4.17.21"
      }
    }
  }' \
  https://api.safedep.io/safedep.services.malysis.v1.MalwareAnalysisService/QueryPackageAnalysis

python

import os
import requests


def check_package(ecosystem: str, name: str, version: str) -> bool:
    response = requests.post(
        'https://api.safedep.io/safedep.services.malysis.v1.MalwareAnalysisService/QueryPackageAnalysis',
        headers={
            'Authorization': os.environ['SAFEDEP_API_KEY'],
            'X-Tenant-ID': os.environ['SAFEDEP_TENANT_ID'],
        },
        json={
            'target': {
                'package_version': {
                    'package': {'ecosystem': ecosystem, 'name': name},
                    'version': version,
                }
            }
        },
        timeout=10,
    )

    if response.status_code >= 500:
        raise RuntimeError('SafeDep unavailable: block install and ask human')

    result = response.json()
    inferred = result.get('report', {}).get('inference', {}).get('isMalware') is True
    verified = result.get('verificationRecord', {}).get('isMalware') is True

    if inferred or verified:
        raise RuntimeError('Malicious package detected: do not install')

    return True

Open Source by Default

Core SafeDep tooling is public. Teams can inspect behaviour and validate integrations independently.

Inspectable Security Decisions

Agent policy is explicit: evaluate inference and verification signals before install, fail closed on uncertainty.

Portable, Not Opaque

Use SafeDep APIs with OSS tools in your own workflows. Reproduce and verify checks independently.

Adopt Safe Defaults
For Agent Installs

Query, evaluate both signals, and fail closed when uncertain. Start free — no credit card required.

Get API Credentials Read API Docs

Stop. Check Every Package.Before You Install It.

Why Agent Dependency Decisions Need Guardrails

Fast Code, Fast Risk

Install Is an Execution Step

One Bad Package Is Enough