There's a question I keep hearing in delivery work that has nothing to do with models:
"Our data lives in the ERP, the CRM, the finance system, and a few spreadsheets. How are we supposed to do AI?"
It's strange. Teams will spend an afternoon debating model choice, then quietly avoid the one thing that actually blocks them — the data doesn't connect.
Helix Foundry (1029 stars, Apache-2.0, TypeScript) is an attempt at that problem.
Connect the tools a company already runs on — databases, Stripe, WorkOS, PostHog, files, generic APIs — and it proposes a cross-source ontology: which invoice belongs to which account, which user holds which subscription.
Then you ask questions over that graph in natural language.
What I find interesting is what it skips: there's no "build a warehouse first" step. The traditional path (ETL → warehouse → BI) is long and expensive, and out of reach for most small teams. This starts by modeling relationships so you can see the shape of your data first.
1. Data stays on your machine by default. No accounts, no hosting, AI runs locally by default. For anyone working under data-compliance constraints (government, healthcare, finance), that default is worth more than any feature.
2. The repo is built for agents to pick up. AGENTS.md and CLAUDE.md both sit at the root, setup logic lives in scripts/setup.sh, and the recommended install is literally "paste this prompt into Claude Code / Codex / Cursor."
That prompt includes a line I appreciated:
Ask me before doing anything that deletes data or stops something that is already running.
Whoever wrote that has done delivery work. The thing you fear most with automation isn't that it fails — it's that it runs off and does something irreversible.
Worth being straight about the limits:
| Item | Status |
|---|---|
| Maturity | alpha |
| Dependency | Docker required; first run downloads several GB |
| Data quality | Does not fix dirty data — you'll just see where it's dirty |
| Deployment | localhost:3001, individual/small-team shape |
It's a POC tool: spend half a day seeing what your data looks like connected, then decide whether to invest further.
I localized the README to Chinese and added a quick-start section on top with the copy-paste agent prompt and Docker notes: https://github.com/yangshun2005/helix-foundry-cn
If it's useful, a star on the original repo helps the authors.