Docs / Install
Install SourceVault
One command on macOS, Linux, and Windows. Every install is free forever on one repository (search, read, Ask), with a 7-day trial of the SourceVault Gateway on that repository. No account, no card, nothing expires.
First, pick a profile
Both profiles are the same install. Lite is what a fresh install runs until the models have arrived, and what it keeps running if you ask for it.
| Lite | Standard | |
|---|---|---|
| Needs | Node. No GPU, no models, no vector database. Any machine with 4 GB of RAM | Node, Ollama, and two local models: about 270 MB for embeddings and 6 GB for the reasoning model, which is the long part of the install |
| RAM | 4 GB | 16 GB, or 8 to 12 GB with the compact model |
| You get | Keyword and symbol-graph search, exact file reads, the dashboard, the Gateway, and the MCP server. Indexing within minutes of the install command | Everything in Lite, plus semantic search and grounded Ask answers |
| Install | The commands below, then sourcevault setup --lite | The commands below as written |
Lite mode is a toggle in the dashboard under Settings → Engine → Lite mode (since v1.55.0). Switched on, SourceVault runs without Ollama; semantic search and Ask decline with a plain message instead of degrading silently. To add the full stack later, switch it off and reindex: the same install upgrades in place. Server deployments (Docker, systemd) set it in the environment instead, and the environment outranks the toggle:
SOURCEVAULT_LITE=1 npm run index-codebase -- <repo-name> SOURCEVAULT_LITE=1 npm start
Lite mode is the fastest way to evaluate SourceVault on a laptop that can't run local models. The Sentinel policy layer, audit chain, licensing, the trial, and the free tier behave identically in both modes.
macOS · Homebrew (recommended)
One command installs everything
Node, Ollama, and SourceVault itself, with config and secrets generated locally at install time. Vectors live in a local SQLite file. There is no database service to run.
# One-time: trust the tap (Homebrew 6+), then install brew trust sourcevault-ai/tap brew install sourcevault-ai/tap/sourcevault # Finish the install: starts the service (and Ollama), pulls the embedding # model and the reasoning model for your RAM (qwen3.5:9b on 16 GB+, # qwen3.5:4b on 8–12 GB) with progress, checks everything, and opens the # dashboard already signed in. Idempotent: run it again any time. sourcevault setup # Put the dashboard in your Dock sourcevault app
Until the models have arrived the dashboard runs lite (keyword and symbol search, file reads, the MCP server) and switches to the full stack on its own, reindexing with vectors. The pieces are commands too: sourcevault open opens the dashboard signed in without ever showing the token (--print writes the one-time link, for ssh), sourcevault doctor checks the install, and sourcevault setup --lite skips the models. The token itself, for another machine or an API client, is grep DASHBOARD_TOKEN "$(brew --prefix)/etc/sourcevault/sourcevault.env".
sourcevault app creates a real macOS app that opens your dashboard, so you get a Dock icon from day one. If you'd rather use the browser, open http://localhost:9000 lands in the same dashboard, where Chrome and Edge show an install icon in the address bar (or Settings → About → Install app) and Safari installs via File → Add to Dock.
From the dashboard, connect GitHub/GitLab/Bitbucket and index your first repository. Or from the terminal:
cd "$(brew --prefix)/opt/sourcevault/libexec" npm run code-repos -- add <git-url> npm run index-codebase -- <repo-name>
Apple Silicon or Intel, macOS 13+. Recommended: 8+ cores, 16 GB+ RAM, 100 GB free disk for the standard profile (qwen3.5:9b, ~6 GB, usable CPU-only). 8–12 GB machines get the compact profile (qwen3.5:4b) automatically; no GPU is required for either. Published benchmark numbers are from the 9b model.
Upgrades: one click in the dashboard (Settings → About → Check for updates) or brew update && brew upgrade sourcevault. Either way the service restarts itself and open tabs reload themselves; there is no manual step. The formula pins each release tarball's sha256.
Linux · Ubuntu / Debian
Pipe the install script to bash
curl -fsSL https://sourcevault.ai/install.sh | bash
To see the plan first, preview without changing anything:
curl -fsSL https://sourcevault.ai/install.sh | bash -s -- --dry-run
What the script does:
- Downloads the latest release and verifies its sha256 and its ed25519 signature against the SourceVault release key before anything is unpacked.
- Installs Node 24 and Ollama (GPU-aware, via Ollama's official installer). Anything already on the box, including an already-running Ollama API, is detected and left alone.
- Installs no vector database and no Docker: vectors live in a local SQLite file. ChromaDB is installed only with
--with-chroma, or kept when an existing install already runs it. - Generates secrets and registers systemd user services.
When the install finishes, sourcevault open opens the dashboard signed in without showing the token (--print for the link over ssh), and sourcevault setup verifies the result end to end and opens it. Both run inside the Linux user that owns the install.
The script is short and readable. Inspect it here before piping it, as you should with any installer.
Updates are one click from the dashboard (Settings → About → Check for updates), or re-run this same command. Either updates SourceVault in place, preserving your index and history. Since v1.54.0 the self-updater downloads the exact release tarball, verifies its sha256 and the ed25519 signature against the key embedded in the build before anything is replaced, and keeps the previous version beside the install with a rollback script (~/.sourcevault/releases/rollback.sh).
Windows · via WSL2 Ubuntu
Start in PowerShell, land in WSL2
SourceVault runs inside WSL2 Ubuntu on Windows. This sets up WSL2 + Ubuntu if needed (first-time WSL installs need one reboot; just re-run the command after), enables systemd, and runs the Linux installer:
irm https://sourcevault.ai/install.ps1 | iex
Before you start, check that WSL2 is allowed on this machine. It needs Windows 10 build 19041 or later, or Windows 11; the Virtual Machine Platform and Windows Subsystem for Linux optional features, which the script enables and which take one reboot the first time; and an administrator account for that one step, after which everything runs as your own user. On a managed laptop those features are often switched off by group policy or by endpoint software. If wsl --install is refused, or your IT department does not permit WSL, nothing in SourceVault will work around it: the installer stops before anything is downloaded.
WSL2 forwards localhost automatically, so the dashboard at http://localhost:9000/dashboard/ works straight from your Windows browser (and installs as a desktop app from there). Updates are one click from the dashboard: Settings → About → Check for updates; since v1.54.0 the release is checksum- and signature-verified before it replaces anything, and the previous version stays available for rollback. Inspect the script.
Managed desktops without WSL: use the team server below instead. One Linux machine or VM runs SourceVault and Ollama; every Windows desktop uses the dashboard in a browser with nothing installed, and the Claude Code and MCP integrations connect to it over the network. For a team on locked-down machines this is the better arrangement anyway: the models run once on a box with the RAM or GPU for them, not on each laptop.
Team server · Docker Compose
One stack, every OS connects via the browser
Run SourceVault on one box (a GPU box for Ollama is optional; ChromaDB only via --profile chroma for deployments that indexed on it); teammates on Windows, macOS, or Linux use the dashboard over the network, installable as a desktop PWA. The optional Qdrant history engine needs no compose entry either: SourceVault manages it inside its own container and state volume when enabled.
Every release since v1.65.0 publishes a container image, ghcr.io/sourcevault-ai/sourcevault, for linux/amd64 and linux/arm64. The release tarball ships the compose.yaml that names it, plus the Dockerfile and the deploy guide. Download and verify the tarball the way the Linux installer does (it, its .sha256 and its .sig sit on the latest release), unpack it on the server, then from inside it:
docker compose pull && docker compose up -d
Since v1.67.0 the same server answers MCP over HTTP at /mcp, so a teammate's Claude Code or Cursor connects to it with a named agent token instead of indexing locally. The MCP page has the client line and what the door enforces.
Set SOURCEVAULT_IMAGE=ghcr.io/sourcevault-ai/sourcevault:v1.65.0 to pin a version instead of following latest. To build from the unpacked source instead of pulling, run docker compose up --build -d. Front the stack with a TLS reverse proxy for production; docs/DEPLOY_SERVER.md in the tarball covers nginx, Caddy, and Traefik.
After the trial
Free tier, or a license
The 7-day trial covers the core product with one source plus the SourceVault Gateway on that repository (access policy, secret redaction, the audit chain, provenance, agent identity), so what you evaluate is the governed path, not just search. When the trial ends, the install becomes a free-tier install (since v1.55.0): search, file reads, and Ask keep working on that one repository on every surface, forever and without a key. Only the Gateway features lock; nothing you indexed is lost, and nothing expires. A license key turns the Gateway back on and raises the source cap: Starter allows three repositories, Pro fifteen. It verifies offline, with no phone-home.
- Pick a plan. Starter and Pro check out instantly, and your key arrives by email.
- Paste the key in Settings → Plan & license and choose Activate.
- Lost your key? Recover it here with your purchase email.