@mocaos/cortex

Interactive installer for self-hosted Cortex — an agentic knowledge base with a chat front end.

npx @mocaos/cortex

Note there is no install subcommand — npx already means "fetch and run", so npx install @mocaos/cortex tries to run a package literally named install and fails with could not determine executable to run.

If you get ENOVERSIONS

npm error code ENOVERSIONS
npm error No versions available for @mocaos/cortex

The package is fine and public. This is npm's supply-chain cooldown: if your .npmrc sets min-release-age (increasingly common after the npm compromises of recent months), npm hides every version published inside that window. A release that is hours old therefore has no eligible versions, and the error says so in a way that sounds like the package doesn't exist.

Check with npm config get min-release-age. To install anyway without weakening the policy anywhere else, override it for the one command:

npx --min-release-age=0 @mocaos/cortex

The flag works on every verb (npx --min-release-age=0 @mocaos/cortex status), and once the release is older than your window plain npx @mocaos/cortex resolves normally. npm has no way to exempt a single package, so it is the flag or the wait — keep the setting; a cooldown is a genuinely good defence, and it is worth more than installing a release on its first day.

What you need

Docker with the Compose v2 plugin (docker compose version). On Linux, apt install docker.io does not include Compose v2; use Docker's official packages or Docker Desktop.

Also curl and tar, which the installer uses to fetch the release artifacts. Both are present on macOS and on virtually every Linux desktop, but Debian's minimal and cloud images ship wget rather than curl — preflight checks for them and names whichever is missing.

npx brings its own Node, but if you already have one on your PATH it must be 20.12 or newer — older Node, including 18, fails the installer's first check before anything is written or pulled.

~20 GB free disk, ~8 GB RAM, linux/amd64 or linux/arm64, and an OpenAI-compatible API key — OpenAI, OpenRouter, Venice, Groq, a local Ollama, or anything else that speaks the same API. Images pull about 1.7 GB total (the backend is the largest piece, at ~1.2 GB).

What it does

Reads the latest tested release manifest, checks your environment, asks what it needs, verifies your LLM credentials with live calls before writing anything, then pulls pinned images and starts the stack.

Two modes: localhost (ports bound to 127.0.0.1) or a public domain with automatic HTTPS via Caddy.

Commands

| Command | Does | |---|---| | npx @mocaos/cortex | Interactive install | | cortex update | Move to the latest tested release | | cortex config | Show where settings live, then restart to apply edits | | cortex status | Service state, health, URLs | | cortex logs [service] | Follow logs | | cortex start / stop / restart | Lifecycle | | cortex backup | Run a verified backup now | | cortex restore [stamp] | Guided restore | | cortex doctor | One pasteable diagnostic block | | cortex uninstall | Remove containers; volumes need a typed confirmation |

Every command accepts --dir <path> if your install isn't ./cortex or discoverable by walking up from the current directory.

Non-interactive

For scripted installs:

[email protected] \
CORTEX_OPENAI_API_KEY=sk-... \
CORTEX_OPENAI_MODEL=gpt-5.2 \
CORTEX_EMBEDDING_MODEL=text-embedding-3-small \
CORTEX_EMBEDDING_DIMENSION=1536 \
npx @mocaos/cortex --yes

Add CORTEX_MODE=domain with CORTEX_APP_DOMAIN, CORTEX_CHAT_DOMAIN and CORTEX_ACME_EMAIL for a public deployment. Secrets are generated unless you supply CORTEX_ADMIN_PASSWORD, CORTEX_NEO4J_PASSWORD, CORTEX_ADMIN_API_KEY, CORTEX_SESSION_SECRET or CORTEX_CHAT_ENCRYPTION_KEY.

Cortex Chat is not installed by default. Add it with:

CORTEX_ENABLE_CHAT=true

In domain mode that also makes CORTEX_CHAT_DOMAIN required. The interactive wizard asks the same question, defaulting to No.

To add or remove chat after installing, edit .env — set or comment out COMPOSE_PROFILES=chat — and run npx @mocaos/cortex restart. In localhost mode that is the only change needed; the chat port and encryption key are already written. In domain mode you also need CHAT_DOMAIN, CHAT_BASE_URL, the chat origin in CORS_ALLOWED_ORIGINS, and cp Caddyfile.chat.template Caddyfile.

Use the installer's restart for this, not a raw docker compose down (with or without --remove-orphans) — Compose filters teardown to the profiles currently active, so once .env no longer selects chat, plain down can't see that container and silently leaves it running. --remove-orphans doesn't help either: it only removes services dropped from the compose file entirely, not one that's merely profile-gated off. restart names the chat profile explicitly so it really is removed; the manual equivalent is docker compose stop chat && docker compose rm -f chat.

Hand-editing COMPOSE_PROFILES=chat yourself works too, in either direction: the next cortex update detects it and corrects cortex.json to match, rather than reverting your edit back to whatever it last knew.

Embeddings default to the chat provider. When they come from somewhere else — Groq serves chat but no embeddings, and Venice is a common embedding pairing — set both of:

CORTEX_EMBEDDING_API_BASE=https://api.venice.ai/api/v1
CORTEX_EMBEDDING_API_KEY=...

Both or neither: a base URL on its own would send the chat provider's key to a different vendor, so it is rejected instead. The wizard asks the same question ("Do embeddings come from that same provider?") and the embedding dimension is always taken from a live probe of whichever endpoint ends up serving them, never assumed.

--yes runs the same live LLM probes as the interactive wizard — before .env exists and before any image is pulled — so a bad key fails in seconds, not after a multi-minute pull.

What it writes

./cortex/
  .env             the only file the installer authors — mode 600, holds your secrets
  cortex.json      install state, no secrets
  docker-compose*.yml, Caddyfile*, ops/, .env.example   release artifacts, verbatim

.env is yours to edit. Compose changes belong in docker-compose.override.yml, which Compose merges automatically and updates never touch.

Privacy

Error reporting is off by default everywhere. The one privacy question the wizard asks — "send anonymous crash reports to the Cortex maintainers?" — only toggles Cortex Chat's server-side reporting; the backend and frontend never report anywhere unless you set your own SENTRY_DSN_BACKEND / SENTRY_DSN_FRONTEND in .env yourself, and neither the frontend nor Cortex Chat ever reports browser-side errors — that half is compiled out of every published image at build time, regardless of what you set at runtime. The installer itself has no telemetry: it talks to GitHub (the release), your registry (image pulls), and the LLM endpoint you configure — nothing else.

Manual install

This installer automates the Compose-based setup documented in selfhost/README.md in the main Cortex repository. Use that path if you'd rather not run npx at all.