What they did
Katyayen set up a Snowflake MCP server exposing one tool of the new CODING_AGENT type, which starts a Cortex agent run in a managed sandbox with bash, file editing and SQL, and connected it to Claude desktop as a custom connector. From a Claude chat they sent a single request: build and deploy a Streamlit dashboard comparing August and September revenue by store, with a 14-day forecast, a store picker and recent customer complaints for a fictional coffee chain's demo tables. The agent explored the tables with read-only SQL, wrote the app and an environment file, ran a test script against the live data, then created a stage, uploaded the files and ran CREATE STREAMLIT under the author's own role. Snowflake asked for approval before each write, and Claude only received a summary and the app link.
How it went
In the final run the app was live in about three minutes after three approvals, with every chart querying the tables live. Two runs produced different apps: one used Snowflake's built-in forecast function, another wrote its own trend model. Earlier attempts broke. The app first failed to load because new apps defaulted to a container runtime, then crashed because the sandbox had a newer charting library than the app runtime. The first build also took more than ten prompts until the instructions told the agent to write, test and deploy in one step each. The author measured roughly 0.4 to 0.6 credits per build.
Worth knowing
The so-called read-only SQL tool still created stages and apps once approved, because the agent acts with the caller's session token, so the role's grants are the real limit. The agent said plainly when it had not opened the app itself, and twice the app had in fact not worked. The data was fictional and the author wrote the post as a lab, so this is one person's tested build rather than a production deployment.