Skip to content
Holotable

Introduction

Holotable turns a natural-language prompt into a live monitoring dashboard.

The one idea to hold onto: the model authors a spec, never data. A panel is a small, validated description of what to compute and how to draw it. Real metric values are only ever produced by the server executing guarded SQL against TimescaleDB — at author time and on every refresh tick thereafter.

Prompt ─▶ /api/generate ─▶ LLM (streamObject) ─▶ Zod IR spec (validated)
                                                     │ save
                                             TimescaleDB (immutable versions)
Viewer ─▶ SSE /stream ─▶ shared in-process poller ─▶ guarded SQL ─▶ TimescaleDB
                                                     └─▶ deltas ─▶ ECharts merge

The stack

Section titled “The stack”

Next.js 16 (App Router) · TypeScript · Tailwind v4 · Base UI · ECharts · TimescaleDB/PostgreSQL for both configuration and metrics · the Vercel AI SDK (streamObject for specs, streamText with tool calls for chat) · Keycloak OIDC with group-based authorization · Server-Sent Events.

The contract

Section titled “The contract”

One shared Zod schema in src/lib/ir.ts is used by the model’s output, the API layer, persistence, and the client. There is exactly one definition, so the contract cannot drift. See The shared IR.

What it deliberately does not do

Section titled “What it deliberately does not do”
  • The model never returns data. It is structurally constrained by the output schema to emit a spec, and its output is re-parsed before it is trusted.
  • The model never chooses a time window. It names a column; the server resolves and injects the bounds.
  • A dashboard never carries credentials. Panels reference sources by an opaque id; credentials resolve from the server environment at execution time.
  • The LLM never runs on view or on a refresh tick. Viewing replays a stored spec.

These are enforced, not aspirational — see Invariants.

Where to go next

Section titled “Where to go next”
  • Quick start — run it with Docker or locally.
  • Your first dashboard — the three steps from an empty install to a live dashboard over your own database.
  • How it works — the path from prompt to live chart.
  • Invariants — the numbered guarantees the design rests on.