Google has announced that Data Agent Kit is generally available. It is a free collection of Model Context Protocol (MCP) tools and Google-authored agent skills that connect more than 15 Google Data Cloud services to coding agents. If you already work with an agent in your IDE or terminal, it can now talk directly to your data on GCP.
What changes
According to Google's weekly Data Cloud roundup (October 5–9), the kit lets developers and data practitioners inspect schemas, author queries and build end-to-end data pipelines in natural language, from the IDE or the CLI. Google lists VS Code, Antigravity, Cursor, Claude Code and Codex as supported environments.
It has two pieces: MCP servers, which expose service operations as tools the agent can call, and skills, which are reusable instructions for specific tasks. The kit is free; the page I read does not detail tool names, installation or usage limits, so read the product documentation before committing to anything.
Why it matters in production
An agent with access to your data warehouse is useful, but it is also one more client with permissions. Three things worry me as an SRE:
- Credential scope: the agent acts with whatever identity you give it. If it is yours, it inherits everything you can see and delete.
- Cost: a natural-language query can scan far more than you expect. Without quotas or per-query limits, an agent can burn money fast.
- Reproducibility: pipelines born in a conversation must end up as versioned, reviewed code, not live in chat history.
The upside is that an open standard like MCP reduces coupling: the same integration works across several agents.
What I would do
- Try it first in a sandbox project with non-sensitive data.
- Create a dedicated service account with read-only roles and, if writes are needed, scoped to a single dataset.
- Set maximum bytes billed per query and budget alerts on the project.
- Treat any generated pipeline or SQL as code: into a repository, with review and tests.
- Check audit logs to see what the agent actually does.
# Example: a read-only service account for the agent (adapt names to your project)
gcloud iam service-accounts create data-agent-ro
gcloud projects add-iam-policy-binding MY_PROJECT \
--member="serviceAccount:data-agent-ro@MY_PROJECT.iam.gserviceaccount.com" \
--role="roles/bigquery.dataViewer"My take: good news for anyone already using agents, as long as it is treated as an integration with permissions and cost, not a toy.