Hosting#

scry is one binary plus one SQLite file. It runs the same everywhere; pick the host that fits and point every device's [client] server_url at it.

Local and home-lab (laptop, mini-PC)#

Bare metal is the simplest install:

cargo install --path crates/scry
cp deploy/scry.service ~/.config/systemd/user/
systemctl --user enable --now scry
loginctl enable-linger $USER    # start at boot, no login needed

Secrets (TAVILY_API_KEY, SCRY_TOKEN) live in ~/.config/scry/env (chmod 600), loaded by the unit's EnvironmentFile and referenced from config.toml as "env:VAR". The config file is then safe to track in dotfiles; the env file never is.

The default config expects an OpenAI-compatible embedding endpoint on localhost:12434 (llama-swap, llama.cpp, Ollama). On a box without one, use the compose bundle instead, which ships its own:

cd deploy
cp config.example.toml config.toml   # edit: embedding base_url = http://llama:8080/v1
SCRY_TOKEN=$(openssl rand -hex 24) docker compose up -d

Reaching a home box from other devices#

Same LAN: set listen = "0.0.0.0:7345", an auth_token, and use http://<box>:7345 as the client server_url.
Away from home: put the box on a WireGuard or Tailscale network and use its overlay address. Nothing about scry changes; it is just a TCP service. This is the recommended setup for a mini-PC.
Public HTTPS: forward a port (or use a VPS as a TCP relay), run Caddy with deploy/Caddyfile.example for TLS, keep the bearer token on.

Cloud / VPS and on-prem#

Same compose bundle. CPU embedding of a 0.6B model is fast enough for queries and incremental syncs; a large first index of a big repo is a one-time cost. Two levers if that bothers you: run the first scry index from a machine with a GPU-backed embedding endpoint, or run the whole stack at home and move the single .db file later; the index carries no host-specific state.

Backups and migration#

The entire index and memory store is one file (db_path). Copy it, and a new host has everything. The embedding model and dimension are stamped inside; a server configured with a different model refuses the file instead of silently mixing vector spaces.

built with cargo soothfast docs build source