proof-of-commitment
Proof of Commitment
Sterne lügen. Verhaltenssignale nicht.
Ein MCP-Server und Web-Tool, das npm-Pakete, PyPI-Pakete und GitHub-Repos nach verhaltensbezogenem Engagement bewertet — Signale, die schwerer zu fälschen sind als Sterne, READMEs oder Download-Zahlen.
Das Problem der Lieferkette
Drei Pakete in einem typischen Node.js-Projekt sind derzeit KRITISCH:
chalk — 399 Mio. Downloads/Woche, 1 Betreuer
zod — 139 Mio. Downloads/Woche, 1 Betreuer
axios — 96 Mio. Downloads/Woche, 1 Betreuer (Angriff am 1. April 2026)
Sterne und README-Qualität zeigen dies nicht auf. Verhaltenssignale schon.
Related MCP server: pkgtruth
Jetzt ausprobieren
Terminal (keine Installation):
npx proof-of-commitment axios zod chalk
# or scan your own project:
npx proof-of-commitment --file package.json
# PyPI too:
npx proof-of-commitment --pypi litellm langchain requestsWeb-Demo (keine Installation): getcommit.dev/audit — fügen Sie Ihre Pakete ein und sehen Sie die Risikobewertungen in Sekunden.
MCP-Server (keine Installation):
{
"mcpServers": {
"proof-of-commitment": {
"type": "streamable-http",
"url": "https://poc-backend.amdal-dev.workers.dev/mcp"
}
}
}Fügen Sie es zu Claude Desktop, Cursor, Windsurf oder einem anderen MCP-kompatiblen KI-Tool hinzu. Fragen Sie dann:
"Überprüfe meine package.json auf Lieferkettenrisiken" "Bewerte axios, zod, chalk, lodash — welches hat das höchste Risiko?" "Wird vercel/ai aktiv gewartet?"
GitHub Action
Fügen Sie Lieferketten-Audits zu jeder CI-Pipeline hinzu — erkennt Pakete automatisch aus package.json oder requirements.txt, postet Ergebnisse als PR-Kommentar, schreibt in die GitHub Step Summary und schlägt optional bei KRITISCHEN Paketen fehl.
# .github/workflows/supply-chain-audit.yml
name: Supply Chain Audit
on: [push, pull_request]
jobs:
audit:
runs-on: ubuntu-latest
permissions:
pull-requests: write # needed for PR comments
steps:
- uses: actions/checkout@v4
- uses: piiiico/proof-of-commitment@main
with:
fail-on-critical: false # set true to block merges
comment-on-pr: true # posts audit table directly on the PRWenn comment-on-pr: true (Standard), postet die Action automatisch die Audit-Tabelle als Kommentar im Pull Request — und aktualisiert denselben Kommentar bei einem erneuten Durchlauf, sodass Sie keinen Kommentar-Spam erhalten. Prüfer sehen die Risikotabelle, ohne den PR zu verlassen.
Eingaben:
Eingabe | Standard | Beschreibung |
| (auto) | Kommagetrennte Paketnamen (automatisch erkannt aus |
|
| Workflow fehlschlagen lassen, wenn KRITISCHE Pakete gefunden werden |
|
| Maximale Anzahl der zu prüfenden Pakete bei automatischer Erkennung |
|
| Audit-Ergebnisse als PR-Kommentar posten (erfordert |
Ausgaben: has-critical, critical-count, audit-summary (Markdown-Tabelle, wird auch in die Step Summary geschrieben).
Beispiel für PR-Kommentar / Step Summary Ausgabe:
| Package | Risk | Score | Maintainers | Downloads/wk | Age |
|---------|-------------|-------|-------------|--------------|-------|
| chalk | 🔴 CRITICAL | 75 | 1 | 380M | 12.7y |
| zod | 🔴 CRITICAL | 83 | 1 | 133M | 6.1y |
| axios | 🔴 CRITICAL | 89 | 1 | 93M | 11.6y |README-Badges
Fügen Sie ein Engagement-Score-Badge zu jedem Paket hinzu, das Sie betreuen oder von dem Sie abhängen:
Beispiele:
Paket | Badge-URL |
axios |
|
zod |
|
litellm |
|
Farben: 🟢 gesund (75+) · 🟡 gut (60–74) · 🟡 moderat (40–59) · 🟠 hohes Risiko (<40) · 🔴 KRITISCH (einzelner Betreuer + >10 Mio. Downloads/Woche)
Badges werden 5 Minuten am Cloudflare-Edge zwischengespeichert. Kein API-Schlüssel erforderlich.
REST-API
Kein API-Schlüssel. Keine Installation.
curl https://poc-backend.amdal-dev.workers.dev/api/audit \
-X POST \
-H "Content-Type: application/json" \
-d '{"packages": ["axios", "zod", "chalk", "lodash", "express"]}'{
"count": 5,
"results": [
{
"name": "chalk",
"ecosystem": "npm",
"score": 75,
"maintainers": 1,
"weeklyDownloads": 398397580,
"ageYears": 12.7,
"trend": "stable",
"riskFlags": ["CRITICAL"]
},
...
]
}7 MCP-Tools
Tool | Beschreibung |
| Batch-Risiko-Audit für bis zu 20 npm/PyPI-Pakete |
| Verhaltensprofil eines einzelnen npm-Pakets |
| Verhaltensprofil eines einzelnen PyPI-Pakets |
| GitHub-Repo-Engagement-Score (Langlebigkeit, Commit-Häufigkeit, Tiefe der Mitwirkenden) |
| Norwegisches Unternehmensregister — Betriebsjahre, Mitarbeiter, Finanzen |
| Dasselbe, nach Organisationsnummer |
| Verhaltensdaten der Browser-Erweiterung (eindeutige verifizierte Besucher, Wiederholungsrate) |
Was der Score misst
Jedes Paket wird auf einer Skala von 0–100 bewertet basierend auf:
Langlebigkeit — Wie lange existiert das Paket schon? Verlassene Pakete werden für Angriffe reaktiviert.
Tiefe der Betreuung — Einzelner Betreuer + Millionen wöchentliche Downloads = die Angriffsfläche, die LiteLLM ausgenutzt hat.
Release-Konsistenz — Regelmäßige Releases signalisieren aktive Überwachung. Lange Pausen = Anhäufung von Schwachstellen.
Download-Trend — Wachsende Pakete ziehen mehr Aufmerksamkeit (und Angriffe) auf sich. Stabil = niedrigeres Profil.
Risiko-Flags:
CRITICAL— einzelner Betreuer + >10 Mio. wöchentliche Downloads (exaktes Angriffsprofil von LiteLLM/axios)HIGH— Paket <1 Jahr alt + schnelle VerbreitungWARN— kein Release in den letzten 12+ Monaten
Echte Datenpunkte
chalk — score 75, 1 maintainer, 399M/week ⚑ CRITICAL
zod — score 83, 1 maintainer, 139M/week ⚑ CRITICAL
axios — score 89, 1 maintainer, 96M/week ⚑ CRITICAL (attacked Apr 1 2026)
lodash — score 88, 3 maintainers, 68M/week
express — score 91, 5 maintainers, 35M/week
litellm — score 74, 1 maintainer ⚑ CRITICAL (supply chain attack Mar 2026)Warum Verhaltenssignale
Der LiteLLM-Angriff (März 2026) und der axios-Angriff (April 2026) folgten demselben Muster: gestohlene Anmeldedaten → bösartiges Paket veröffentlicht → 97 Mio.+ Maschinen exponiert. Beide Pakete wurden vor den Angriffen durch diese Metriken als KRITISCH eingestuft.
Deklarative Signale (Sterne, README-Qualität, CI-Badges) erfassen dieses Risiko nicht. Verhaltensbezogenes Engagement schon.
Im offiziellen MCP-Register gelistet
registry.modelcontextprotocol.io → io.github.piiiico/proof-of-commitmentStack
Ebene | Technologie |
Backend | Cloudflare Workers + D1 |
MCP | Model Context Protocol SDK |
Daten | npm registry, PyPI, GitHub API, Brønnøysund (NO) |
Landing | Astro + Cloudflare Pages |
Die umfassendere Vision
Lieferketten-Audits sind das erste Werkzeug. Das zugrunde liegende Primitiv ist ein Engagement-Graph — Verhaltenssignale, die inhaltsbasiertes Vertrauen über jede Domäne hinweg ersetzen.
Wenn Inhalte leicht zu fälschen sind (Bewertungen, Sterne, READMEs), wird Engagement zum Signal. Ein Betreuer, der in 12 Jahren 847 Releases veröffentlicht hat, zeigt eine andere Art von Engagement als jemand, der 2023 einmal etwas veröffentlicht hat.
Die gleiche Logik gilt für Websites, Unternehmen und KI-Agenten. Zwei Kartennetzwerke haben diese Lücke unabhängig voneinander benannt: Mastercard Verifiable Intent §9.2 listet verhaltensbasiertes Vertrauen explizit als "nicht abgedeckt". Visa TAP identifiziert Agenten, ohne zu beantworten, ob man ihnen vertrauen sollte.
Proof of Commitment ist die Vertrauensebene, auf die sie hinweisen.
Lokal ausführen
bun install
bun run dev:backend # local server with SQLite
bun run test:e2e # E2E test with mock World IDBereitstellen:
bun run deploy # deploys to Cloudflare WorkersAvailable Tools
8 toolsaudit_dependenciesAInspect
Batch-score multiple npm or PyPI packages for supply chain risk. Takes a list of package names and returns a risk table sorted by commitment score (lowest = highest risk first).
Risk flags:
CRITICAL: single npm publisher + >10M weekly downloads (publish-access concentration risk)
HIGH: new package (<1yr) + high downloads (unproven, rapid adoption = supply chain risk)
WARN: no release in 12+ months (potential abandonware)
WARN: dormant publishers with current scope access — contributors who stopped publishing but retain npm tokens (the Mastra incident vector, June 2026)
Perfect for auditing a full package.json or requirements.txt — paste your dependency list and get a prioritized risk report.
Examples: score all deps in a project, compare two similar packages, identify abandonware before it becomes a CVE.
Inputs: packages (required), ecosystem.
lookup_businessBInspect
Search for a Norwegian business and get its commitment profile from public data (Brønnøysund Register Centre). Returns real commitment signals that can't be faked:
Temporal commitment: how long the business has operated
Financial commitment: revenue, profitability, equity health
Operational commitment: employee count, active status
Overall commitment score (0-100)
Data source: Norwegian government registers (Brreg). No user-contributed data needed — immediate trust verification for any Norwegian business.
Inputs: query (required), maxResults.
lookup_business_by_orgAInspect
Look up a specific Norwegian business by organization number and get its commitment profile from public data (Brønnøysund Register Centre). Returns real commitment signals: longevity, financial health, operational activity, and overall commitment score.
Inputs: orgNumber (required).
lookup_github_repoAInspect
Get a behavioral commitment profile for any public GitHub repository. Returns real signals: how long the project has existed, recent commit frequency, contributor community size, release cadence, and social proof. These are behavioral commitments — harder to fake than README claims.
Useful for: vetting open-source dependencies, evaluating AI tools/frameworks, assessing vendor reliability. Examples: "vercel/next.js", "facebook/react", "https://github.com/piiiico/proof-of-commitment"
Inputs: repo (required).
lookup_go_moduleAInspect
Get a behavioral commitment profile for any Go module on proxy.golang.org. Takes a full module path (e.g., "github.com/gin-gonic/gin", "golang.org/x/net", "k8s.io/client-go", "gopkg.in/yaml.v3") and returns real signals: module age, version count, publish cadence, GitHub contributors (the closest equivalent to "publishers" since Go has no centralized publisher concept — git push access is the publish equivalent), GitHub stars, OpenSSF Scorecard score.
The Go ecosystem has no centralized download counter, so this profile is GitHub-primary — the linked source repository's activity, contributor count, and Scorecard carry more weight than for npm/PyPI/Cargo. Stars are used as the popularity proxy.
Useful for: vetting Go dependencies before adding to go.mod, identifying abandonware, supply chain risk assessment. Examples: "github.com/gin-gonic/gin", "golang.org/x/crypto", "github.com/spf13/cobra", "k8s.io/api"
Inputs: module (required).
lookup_npm_packageAInspect
Get a behavioral commitment profile for any npm package. Returns real signals: package age, download volume and trend (growing/stable/declining), release consistency, npm publisher count, GitHub contributor count, and linked GitHub activity.
Also returns publisherLifecycle — cross-referencing current maintainers against per-version publish history to flag dormant publishers who still hold npm scope access. The Mastra incident (June 2026) exploited exactly this: a contributor dormant since 2024 with never-revoked scope access.
Supply chain attacks target packages with low publisher depth (few people with npm publish access). Behavioral signals reveal what download counts hide.
Useful for: vetting dependencies, identifying abandonware, due diligence on open-source packages. Examples: "langchain", "@anthropic-ai/sdk", "express", "litellm"
Inputs: package (required).
lookup_pypi_packageAInspect
Get a behavioral commitment profile for any PyPI (Python) package. Returns real signals: package age, download volume and trend, release consistency, publisher/owner count, and linked GitHub activity.
Supply chain attacks target Python packages — LiteLLM (97M downloads/mo) was compromised via stolen PyPI token in March 2026. Behavioral signals reveal what star counts hide.
Useful for: vetting Python dependencies, identifying abandonware, supply chain risk due diligence. Examples: "langchain", "litellm", "openai", "anthropic", "requests", "fastapi", "pydantic"
Inputs: package (required).
query_commitmentAInspect
Query verified behavioral commitment data for a domain. Returns aggregated signals: unique verified visitors, repeat visit rate, and average time spent. These prove real human engagement — harder to fake than reviews or content.
Inputs: domain (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 8 tools
The package-ecosystem lookups (npm, PyPI, Go, GitHub) are clearly scoped to distinct registries, and audit_dependencies is obviously batch-oriented. The only real overlap is lookup_business vs lookup_business_by_org, which differ mainly by input (name search vs org number), and query_commitment's domain-based scope is somewhat fuzzy against the package lookups since package profiles also aggregate engagement signals.
All eight tools use snake_case verb_noun form, with six sharing a lookup_ prefix and the rest using query_/audit_ verbs appropriately. Minor inconsistency in the noun portion of the lookup family (github_repo vs go_module vs pypi_package vs npm_package vs business), but the pattern remains predictable.
Eight tools is well inside the sweet spot for a registry/commitment-scoring server. Each tool maps to a distinct data source or aggregation mode, and none feels redundant padding.
The surface covers Norwegian business data, four code-ecosystem sources, domain engagement, and batch risk auditing — a coherent lifecycle from single lookup to prioritized audit. Gaps remain: audit_dependencies only handles npm/PyPI (not Go/GitHub), and other major ecosystems (Cargo, RubyGems, Maven) have no lookup at all, but agents can work around these.
Maintenance
Related MCP Connectors
Audit GitHub repos for malicious and supply-chain code before you depend on them.
Package versions, licences and known vulnerabilities for npm, PyPI, crates, Go, Maven, NuGet
Evaluates any GitHub repository for license risk, maintenance velocity, stars/forks, and supply-c.
Safety grade (A to F) for any public GitHub repo, in plain English, before you install it.
Related MCP Servers
AlicenseNot gradedqualityBmaintenanceDeterministic supply-chain provenance, SBOM/AI-BOM generation, and dependency risk analysis for npm and PyPI packages, with 14 security rules and verifiable reports.1MIT- AlicenseAqualityBmaintenanceEnables AI coding agents and CI to vet npm dependencies before they reach the lockfile, flagging hallucinated, slopsquatted, or otherwise risky packages with evidence-backed verdicts.352 npm4MIT
- AlicenseNot gradedqualityBmaintenanceEnables safety-ranked discovery of open-source components across npm, PyPI, crates.io, RubyGems, GitHub, and Hugging Face, with license, security, and health checks.1,348 npmMIT
- AlicenseAqualityBmaintenanceScores npm packages, PyPI packages and GitHub repositories 0-100 with an A-F grade and plain-language risks. Scans package.json, requirements.txt and lock files; no API key needed.367 npm1Apache 2.0