Skip to main content
metalworks does not prescribe a database. Storage is a set of typed repository protocols (CorpusRepo, BriefRepo, RunRepo, AccountRepo, OpportunityRepo, InboxRepo). One backend object implements as many as it supports. Two backends ship in core, with zero infrastructure: 
from metalworks.stores import MemoryStores, SqliteStores

mw = Metalworks(store=MemoryStores())                         # ephemeral, the default
mw = Metalworks(store=SqliteStores("~/.metalworks/store.db")) # persistent, WAL-mode
MemoryStores and SqliteStores each satisfy every repo, so either drops into Metalworks(store=...), ResearchDeps, or the Reddit OAuth account store.

​
Your own backend

Implement the repo methods you need on your own class and pass it as store. Verify it against the same conformance suite the built-ins pass — including the
1000-rows-behind-one-filter pagination case that catches silent truncation: 
from metalworks.testing import check_all_repos

check_all_repos(MyStore(), corpus_rows=1500)
Anything that exposes the repo methods works: Postgres, DynamoDB, a REST service, an in-process dict. There is no generic document store to bind — the repos are typed because real production tables are columnar.

​
Supabase / PostgREST

A Supabase backend is one such custom store. If you want one, implement the repos over the Supabase client (paginating reads with .range() to exhaustion, since PostgREST silently caps result sets at its max-rows setting and returns HTTP 200). Bind it to pre-existing columnar tables by mapping logical collections to your physical table names in the constructor. This is a backend you write for your own infrastructure, not something metalworks configures for you.

​
Pick the right one

BackendUse it forInfra
MemoryStorestests, one-shot runs, the offline demonone
SqliteStoresa single machine, persistence across runsnone
your owna service, a shared DB, multi-tenantyours