A 3-person data team replaced a $150k Looker licence with Claude-written HTML dashboards

From afternoon prototype to a sandboxed platform

David Lequin's post describes a 3-person data platform team at a roughly 500-person company. A colleague's February prototype hosted Claude-generated HTML in S3 behind SSO; users liked pages that looked better than Looker's fixed tiles, and weekly viewers rose from 23 in March to 136 in June. The first version had a serious flaw: dashboard JavaScript ran on the same origin as the viewer's session.

The rebuild runs dashboards in sandboxed iframes on their own origin, with no cookies and a strict CSP. Snowflake handles identity: queries run with the viewer's own token, so sharing a dashboard never grants data access. The bill: about $6.4k of Claude tokens, $1.9k of AWS so far, and roughly $6k a year to run.

Where the thread pushed back

A co-founder of skipup.ai quoted the post's line that a picture now costs about $0.50 of tokens and called it "an awfully simplified view of BI": most companies struggle to define what to measure, and analysts print pictures while they work that out.

A Retool employee, whose company says "probably the majority" of its business is replacing SaaS with citizen-developed apps, agreed the category will shrink but named the bottleneck: data quality, with Snowflake and Databricks schemas of 10^5 tables where "just putting a LLM on that doesn't work."

The number the author left out

One commenter said counting only tokens ignores the salaries of the team that built and maintains it, and guessed the saving is "more like 50%." The post itself concedes that engineering hours are not counted and that the approach only works for a team with a mature Snowflake and dbt stack. The takeaway for readers: the licence you can drop is the visualization layer, while the metric definitions are what you still pay for.