Install in one command

Quickstart

The fastest path to a running brain is the one-command installer. It takes a clean machine that has only Docker and Node, pulls prebuilt images, writes a hardened configuration, boots the stack, pulls the local inference models, and prints the console URL. There is no source clone and no local image build.

flowchart TD
  accTitle: Setup flow
  accDescr: Run the installer, copy the setup token, open the setup page and paste it, create the first administrator, confirm inference, connect a source, and arrive at a queryable company brain.
  R["npx @connectai/selfhost run"] --> T["npx @connectai/selfhost token"]
  T --> SU["Open localhost:5273/setup<br/>paste the token"]
  SU --> AD["Create the first administrator"]
  AD --> INF["Confirm inference"]
  INF --> C["Connect a source"]
  C --> DONE["Queryable company brain"]

Prerequisites

Docker with the Compose v2 plugin, and the daemon running.
Node 18 or newer (for npx).
About 8 GB of RAM and 10 GB of free disk for the images plus the local models. Less than that is an advisory warning, not a hard stop.

That is the whole list. No git, no clone, no source checkout.

Run it

bash

npx @connectai/selfhost run

The boot is quiet by default: a spinner with a short summary, not raw per-layer docker output. Add --verbose (or set CONNECTAI_VERBOSE=1) for the full stream.

The quiet boot console of npx @connectai/selfhost run

If you would rather install the command globally:

bash

npm i -g @connectai/selfhost
connectai run

What the installer does

Pre-flight. Verifies a running Docker daemon and Compose v2. It warns (and does not block) if RAM or disk is below the footprint. Errors are copy-pasteable.
Materialize assets. Writes the bundled deploy files into a working directory (default ~/.connectai-selfhost, override with --dir). You never fetch source.
Generate `.env`. Fills the secrets with a cryptographic random generator (ENCRYPTION_SECRET, the Infisical vault identity, and a strong POSTGRES_PASSWORD), writes the file chmod 600, and never overwrites a value you already set. Secrets stay local: never logged, never transmitted. A second run is a no-op.
Boot. Pulls the prebuilt images, brings up the bundled Infisical vault, provisions the vault identity, brings up the full stack, and waits for the API /health check and the console.
Models. Pulls qwen2.5:1.5b (chat) and mxbai-embed-large (embeddings) into the bundled Ollama, and gates the loop worker so ingest never runs before the embedder is ready.
Print. Shows the console URL, the setup-token command, and your single next step.

The first run downloads several GB (images plus models) and can take a few minutes. Re-runs are fast and idempotent.

When it finishes

Console: http://localhost:5273
API health: http://localhost:4000/health

Get the one-time first-run setup token (printed once at API boot) with:

bash

npx @connectai/selfhost token

Then open http://localhost:5273/setup, paste the token, create the first administrator, confirm inference, and connect a source.

Commands and flags

Command	What it does
`run` (default)	the full install and boot above
`token`	print the one-time first-run setup token
`logs`	tail stack logs (for example `connectai logs api`)
`down`	stop the stack (`down -v` also removes volumes and wipes the brain)
`help` / `version`	usage / version

Flag	Meaning
`--dir <path>`	working directory (default `~/.connectai-selfhost`)
`--verbose`	stream the full docker output during boot (default: quiet spinner)
`--yes`, `-y`	non-interactive (already the default; accepted for CI)
`--tag <tag>`	run a specific release tag (advanced; also `CONNECTAI_IMAGE_TAG`)

Notes

localhost is the v1 target. The prebuilt console talks to the API at http://localhost:4000. The compose ships a runtime API-base seam, so serving on a remote origin is a configuration change rather than a rebuild. Full remote serving is a documented follow-up.
The installer is a thin front door over the same hardened boot engine and canonical compose used everywhere else. It does not fork the orchestration.

Where to go next

Connect your agent (MCP) points Claude, Cursor, or a CLI at your new brain.
Self-hosting (advanced) covers the manual docker-compose path if you need to tune environment, vault, or models beyond the installer defaults.