Data Agent Kit: Run cross-system data investigations from your IDE

What the Data Agent Kit is

The kit is packaged so developers don't need to leave their editor to answer open-ended data questions that normally require hopping between consoles and tabs.

How it changes routine investigations

A typical investigation often touches analytical stores, operational databases, and raw metadata files. The example used in the original coverage shows a single chat-driven workflow that:

• Drills into January's AOV by order type and finds a new channel with lower AOV.
• Queries the Cloud SQL Postgres customer table to identify that the new channel's accounts are recently created business accounts.
• Reads campaign JSON files in Cloud Storage to discover a 25% discount promo code driving many low-value wholesale orders.

This sequence demonstrates the kit's intended value: an agent can execute the relevant queries and follow leads across systems in one conversation instead of copying SQL between consoles.

Permissions and auditability

Before the agent executes queries, the IDE pauses for explicit permission to use MCP tools (for example, an execute_sql_readonly call). Users can allow a single run (for audit logging) or choose "always allow" for repeated workflows. The IDE exposes the execution trail, including each MCP tool call and the raw SQL sent to the backend, so developers can review generated code and audit what ran.

Where it runs and how you use it

Cursor, and as plugins for other tools such as Antigravity 2.0, Antigravity CLI, Claude Code, and Codex. Skills are authored as markdown files that augment the agent's understanding of your environment and how to interact with specific services.

Practical tradeoffs to consider

• Permissions model is explicit, which helps auditing but requires attention to avoid frequent prompts during iterative work.

Bottom line

If you regularly trace hard-to-pinpoint metric changes across analytics, operational, and file stores, the Data Agent Kit packages a workflow that keeps the investigation inside your IDE. It routes agent calls through MCP servers, relies on authored skills for environment knowledge, and surfaces an execution trail so developers can approve and audit automated queries.