Schemas
live · betaWrite DBML, get a diagram you can actually read. Zoom from context to columns with C4 levels, link elements, and discuss them in comment threads right on the table.
- DBML
- C4 levels
- links
- comments
- live-DB drift · planned
Data models, APIs, flows and decisions. Kept current, linked together, and readable by your team and by the agents working next to them.
01 the gap
developers
“Which tables does this touch? Is the diagram still true?”
product managers
“The spec says one thing. The build does another.”
QA
“Which edge cases exist? Who makes the test data?”
AI agents
“No shared context, so we guess. Confidently.”
02 modules
Write DBML, get a diagram you can actually read. Zoom from context to columns with C4 levels, link elements, and discuss them in comment threads right on the table.
OpenAPI, AsyncAPI and GraphQL, linked to the tables they read and write. Find breaking changes before prod does.
User flows and state machines, linked to endpoints and tables, so specs stop drifting from the build.
Change something, see the flows, endpoints and tests it affects. Generate fixtures that respect every foreign key.
Architecture decisions pinned to the elements they explain, so the “why” stays when people move on.
03 zero-access security
Your schema is a map of your business. End-to-end encryption means the map stays yours, even from us.
in progress End-to-end encryption is rolling out in the Schemas beta. Until it reaches your workspace, content is protected by TLS and access control, not by your keys.
Keys are created in your browser. The server stores only ciphertext (XChaCha20-Poly1305, X25519 key wrapping).
DBML, layout, titles, notes and comments are encrypted. We see what’s needed to run the service: ids, members, roles, sizes, timestamps.
Servers at Hetzner in Germany. No trackers and only essential cookies, so no cookie banner.
Export everything, delete your account for real, backups purged within 30 days.
04 built for AI agents
Claude Code, Cursor and friends ask “how are orders related to users?” or “what breaks if I drop x?”. Agents propose changes as reviewable suggestions, never silent edits.
One link with the schema, APIs, flows and decisions for a scope. Hand it to a contractor, an integrator, or an AI.
Publish a public card with your APIs and data model, so other companies’ agents can find you and integrate.
Fits zero-access: an agent reads through a local Beintalk connector that holds your keys. The server still sees nothing.
$ curl -s https://beintalk.com/api/ai/challenge { "id": "ch_5f0c…", "expiresAt": "+20 s", "schema": "Table orders {\n id int [pk] …", // 8–12 tables "question": "List every table that references `users` directly or transitively … sorted JSON array", "pow": { "algo": "sha256", "difficulty": 5 } } $ node ai-subscribe-demo.mjs --operator "Acme Labs" \ --contact ops@acme.example --topics schemas,ai ✓ parsed 11 tables, 14 refs 3 ms ✓ answer ["carts","order_items","orders","reviews"] ✓ proof-of-work nonce=1kq2t 612 ms ✓ subscribed sub_9c4e… → poll /feed.json?topics=ai,schemas
AIs can subscribe to our news, but they have to prove they’re AIs. Fetch a random schema, answer a question about its foreign keys, add a small proof-of-work, all in about 20 seconds. Easy for a program, tedious for a human with a browser.
Honest fine print: this proves “an automated agent that can reason and run code”, not a cryptographic AI identity. A human with a script passes too. We’re fine with that.
05 plans
solo
Free
team
Per seat pricing coming
When a trial ends or you downgrade, extra diagrams become read-only. Never deleted.
06 newsletter for humans
Launches and the odd lesson learned. Double opt-in, one-click unsubscribe. Robots, there’s a better door.