DB Branching for AI Agents

Give AI agents a table branch before they write, review then merge

Problem

The poster, working on agent deployment with MatrixOne, found that letting an agent read data is easy but letting it write is scary. Telling the agent in the prompt 'don't mess with production data' doesn't work — once it has write access, a single overly broad WHERE clause can silently change thousands of rows, and afterward you can't even tell which rows changed. This is a concrete, high-frequency, and dangerous scenario.

Solution

A lightweight branching layer for existing databases: connect to any MySQL/Postgres database, create table-level snapshots and writable branches for agents, route all agent writes to the branch, provide row-level diff, and merge back to the main table after approval with conflict policies. Available as CLI or MCP service for agent workflows.

Users

Data engineers, backend developers, and data platform teams using AI agents on databases; DBAs needing safe audit trails

MVP

Connect to MySQL/Postgres; create snapshot branches for selected tables; route agent writes to the branch; row-level diff between branch and main; merge or discard after approval

Revenue

Open-source self-hosted + paid cloud (managed branch storage, audit logs, team permissions); or per-connection/per-database subscription

Why now

AI agents are moving from read-only chat to write operations, yet databases lack a safety isolation layer designed for agents. Building a branch proxy on top of mature databases is far lighter than building a database, and fits an indie developer.