Eatmydata.ai brings local-first question-to-SQL dashboards to the browser
Eatmydata.ai is a fully frontend, open-source app where AI agents generate SQL queries and dashboard code against in-browser SQLite — with data never leaving the browser.
Score breakdown
The tool demonstrates a fully local-first agentic data-analysis workflow where the remote LLM never accesses raw data, addressing both privacy concerns and the performance limitations the author observed with large datasets in general-purpose AI chat tools.
- 01Fully frontend app — data never leaves the browser; the only remote call is to an LLM via OpenRouter or similar provider.
- 02An Orchestrator → SQL Planner → Coder agent loop generates SQL queries and dashboard configuration code from natural-language questions.
- 03All data sent to the remote LLM is heavily sanitized and obfuscated; the LLM never sees actual data contents.
Eatmydata.ai is a fully frontend application — with the sole exception of a remote LLM call via OpenRouter or another provider — that allows users to query their data using natural language and receive generated dashboards without any data leaving the browser. The core agent loop consists of an Orchestrator, SQL Planner, and Coder agent that collaboratively produce multiple SQLite queries and dashboard configuration code. Queries run against an in-browser SQLite instance using the OPFS adapter from `wa-sqlite`, and generated code executes inside a sandboxed QuickJS environment. Dashboards are rendered using Apache ECharts.
Semantic indexing is also fully local, using quantized embedding generation models and a `TurboQuant` semantic indexing extension for SQLite running entirely in the browser.
Privacy is a central design goal: all data sent to the remote LLM is heavily sanitized and obfuscated at several points, so the LLM never sees the actual contents of the data it analyzes. Semantic indexing is also fully local, using quantized embedding generation models and a `TurboQuant` semantic indexing extension for SQLite running entirely in the browser. The project's custom NER and embeddings inference engines are written in zero-dependency C with `wasm-simd128` optimizations, which the author describes as 1.7x faster and 38x lighter in binary size compared to `onnxruntime`. The tool also includes a fork of the xlsx Community Edition with style support, and is MIT-licensed with source available at `github.com/eatmydata-org/eatmydata`.
Key facts
- 01Fully frontend app — data never leaves the browser; the only remote call is to an LLM via OpenRouter or similar provider.
- 02An Orchestrator → SQL Planner → Coder agent loop generates SQL queries and dashboard configuration code from natural-language questions.
- 03All data sent to the remote LLM is heavily sanitized and obfuscated; the LLM never sees actual data contents.
- 04Semantic indexing (embeddings + vector search) runs entirely locally using a TurboQuant SQLite extension.
- 05Custom NER and embeddings inference engines use zero-dependency C and wasm-simd128 optimizations, described as 1.7x faster and 38x lighter than onnxruntime.
- 06AI-generated code runs inside a sandboxed QuickJS environment; dashboards are rendered with Apache ECharts.
- 07Released under the MIT license as a static web app at github.com/eatmydata-org/eatmydata.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 10, 2026 · 15:34 UTC. How this works →