What it does

Lealone is a Java-based database engine that doubles as an AI agent. You start it with a -agent flag, configure an LLM (Doubao or DeepSeek), then type natural-language requests like “implement a todo app.” It returns a URL. You can also define services, tables, and workflows in a SQL-like syntax and execute them immediately, or point it at a .sql file to bootstrap an entire application.

The interesting bit

The project treats SQL as both schema language and orchestration layer for LLM-driven code generation. A create workflow statement can include a natural-language comment that presumably drives agent behavior — the README shows a workflow whose comment says “find the specified user, greet them, tell them the current time.” The database, the services layer, and the AI agent all live in one JAR.

Key highlights

• Single JAR deployment (java -jar lealone-8.0.0-SNAPSHOT.jar)
• Built-in HTTP service endpoints auto-generated from create service definitions
• Supports ACID transactions, replication, and sharding per its topic tags
• LLM configuration happens through SQL: set llm (provider: '...', model: '...', api_key: '...')
• “Atmosphere programming” (氛围编程) — describe an app, wait ~20 seconds, get a URL

Caveats

• README is entirely in Chinese; English documentation appears absent
• Only two LLM providers supported currently (Doubao, DeepSeek)
• Claims to “thoroughly subvert existing software development models” — the actual mechanism for this subversion is left as an exercise to the reader
• Version is 8.0.0-SNAPSHOT; stability and production readiness are unclear

Verdict

Worth a look if you’re building Chinese-market enterprise tools and want to experiment with LLM-generated CRUD apps inside a database runtime. Pass if you need mature docs, broad LLM support, or a clear boundary between your data layer and your AI orchestration.