Run Upwarden inside your perimeter.

1 · What you need

Three datastores and one outbound dependency — all stable, mature components you may already run:

• PostgreSQL 17+ with pgvector — audit log, policy, scanner verdicts and the embedding index. pgvector is required (the scanner does similarity search over package-content embeddings).
• Redis 7+ — rate limiting, manifest cache, idempotency and fast-pass coordination. Cluster mode supported.
• Object storage — tarball staging for the scanner. GCS by default; S3/MinIO-compatible for portability.
• An embedding endpoint — any OpenAI-compatible API for the similarity scanner's vectors.

A Stytch account is optional — only the hosted dashboard auth needs it; operators running with the admin-token bearer can skip it.

2 · Run the container

Upwarden ships as a single image. The simplest path is docker compose — the compose file boots the engine plus a local Postgres and Redis for evaluation (swap in your managed datastores for production). Migrations run automatically on boot.

$ git clone https://github.com/upwarden-io/upwarden.git
$ cd upwarden
$ cp .env.example .env
# set DATABASE_URL, REDIS_URL, GCS_BUCKET, EMBEDDING_ENDPOINT, …
$ docker compose up -d

For Kubernetes, the manifests in k8s/ are the reference — the same ones we run in production for Upwarden Cloud. Set SCANNER_ENABLED=true and the datastore URLs; every variable is in the environment reference.

3 · Create your first key

Each install authenticates with a vk_… project key. As admin, create a tenant, a project, and a key from the dashboard or the JSON admin API. Full walkthrough: Create your first key.

4 · Point your package manager

Exactly like the managed service — one registry line per ecosystem (npm, bun, yarn, pnpm, pip, uv, cargo, Go, Maven, NuGet, RubyGems), pointed at your instance hostname instead of ours. The Quickstart covers each ecosystem in two lines.

5 · Verify it works

Confirm the engine is up, then prove the block path:

$ curl -fsS https://npm.pkg.your-domain.com/healthz
{"ok":true}

Then install a known-blocked version and watch it roll back. The full three-step proof is in Verify it's working.

What stays yours

• Nothing leaves your perimeter. No package names, machine identifiers or org names egress — zero telemetry, by design, not a setting.
• You hold the audit log. Every manifest served, block, quarantine and rewrite is recorded in your database, tied to client identity.
• Explainable verdicts. Every decision carries a reason your team can audit and defend.
• Going further: SIEM export and audit streaming (gzipped JSONL) are available today at Org+; fully air-gapped operation is on the enterprise roadmap.