KaiwaDB
DocsOpen dashboard

Ask your database
in plain English.

KaiwaDB compiles a plain-English question into a standalone query that runs on your own database, returning the answer and a diagram of how it was derived. No warehouse migration. No shipping rows to a cloud model.

Open dashboard →Read the docs
SCHEMAS / speedlo
How many orders in the past 30 days?Generate ↗
QueryStatus
Orders in past 30 daysApproved
Orders per branchPending
Result · 30d
14,203 orders
D2 · flow
list orders
filter last 30d
count rows
One query, compiled for your database engine
Two ways in

Pick your tradeoff: effortless onboarding, or maximum privacy.

Same natural-language interface. You decide how much KaiwaDB touches.

↔
Tunnel
Auto-connect

A lightweight tunnel deployed next to your database connects outboundto KaiwaDB. It reads your schema automatically, so there's no manual setup, ask a question and results come straight back.

  • ✓Automatic schema introspection, zero-effort onboarding
  • ✓Queries auto-run through the tunnel
  • iIn this mode KaiwaDB is in the data path and sees queries and rows.
⚿
API-only
Privacy path

Keep KaiwaDB out of the data path entirely. You send only your schema and questions; we return the compiled query. You run it yourself.

  • ✓Rows and credentials never reach us
  • ✓Only schema and questions cross the perimeter
  • ✓You stay fully in control of what runs
How it works

Best of both worlds: a smart model writes, a private model runs.

Frontier models are smart but hungry for compute, so it is rarely worth dedicating that much VRAM to private, in house use. Models small enough to self host are not smart enough to write good queries. KaiwaDB splits the work so each model does only what it is good at.

01
Frontier model
Smart enough for the hard part: writing a clean Kaiwa query from your schema and aliases. Never sees a row.
02
Standalone query
Compiled to your SQL dialect, a self-contained artifact, approvable and reusable.
03
Runs on your DB
The query executes where your data already lives. Run it once, or a million times.
04
Private model rephrases
A small model you can self host does the easy part: turning the raw result into plain language, inside your infra.
The metaphorThe AI writes the recipe once, cooking it is free.
Explainability

One question. The data, and the diagram of how you got it.

Because a single Kaiwa query compiles to both a result and a D2 flow diagram, every answer can ship with a picture of how it was assembled, which tables, filters, joins and aggregations were involved. Tools that just poke your database and hand back a number can't do that.

Orders in past 30 daysRESULT
Result
14,203
orders · +18% vs prior 30d
Why KaiwaDB

Built as a compiler, not a prompt template.

01
One compile, infinite runs
A question becomes a standalone artifact once. After approval it runs a million times at near-zero marginal AI cost.
02
Streams, not just chat
Point a standalone query at a change stream and let it keep producing answers as new data lands.
03
No platform migration
KaiwaDB sits in front of the Postgres, ClickHouse or BigQuery you already run. Nothing to move.
04
Multi-target via one IR
A shared intermediate representation renders to SQL dialects, Mongo pipelines (WIP), and D2 diagrams.
05
Analyst-in-the-loop
Humans approve queries before they become reusable, and every approval teaches your company terminology.
06
Meet users where they are
Ask from the product, the API, or a first-class Telegram bot, no separate app required.
Pricing
Coming soon

A subscription, plus tokens for pipelines.

Think Claude Code + API credits: a base plan for the product, and metered tokens for the frontier-model work of generating pipelines. Once a pipeline is approved, running it is effectively free.

Plan
Subscription
TBD/ month
The product surface: schemas, aliases, analyst review, Telegram bot, streams and dashboards.
Metered
Pipeline tokens
TBD/ token pack
Frontier-model generation is billed per token. Buy packs as you go, approved pipelines then run at near-zero marginal cost.
Final pricing is being finalized, talk to us for early access terms.

Point us at your database.

Managers ask in plain English, analysts stay in control, and your data stays exactly where it belongs.

Open dashboard →Read the docs