Can we replace a $150k Looker with a bucket of HTML files?
B · Useful, limited First-hand account from data platform engineer David Lequin with specific build and run costs and a concrete security design; useful practice, but a single company's self-reported figures with engineering time uncosted.
Summary based on the full text·Original: about 11 minutes to read
In short
A blog post (14 October 2026) by David Lequin, one of three data platform engineers at a company of about 500 people. What began as an afternoon side project to host Claude-generated HTML dashboards became an internal BI tool used as much as Looker, whose use fell by a third. The build cost about $6.4k in tokens; running costs about $20k a year. His argument: Snowflake already held permissions and dbt the semantics, so the part of Looker left to replace was turning queries into pictures.
Key numbers
- ~$6.4kClaude tokens spent building the dashboard platform over 8 months (325 commits, ~74,000 lines)
- ~$20k a yearrunning and extension cost, against a $150k Looker licence
- $2.50 to $0.50token cost per dashboard built by the in-house analyst agent, before vs after a model switch
- 23 to 136weekly dashboard viewers, March to June
As reported by the source.
English·Web page·opens david.lequin.dev
What it reports
- February: an S3-backed service behind SSO, with a script packaged as a Claude plugin to publish HTML. March: a live query API using the viewer's Snowflake OAuth token, so data access follows existing Snowflake roles.
- Weekly viewers rose from 23 in March to 136 in June; executives began circulating links.
- The first version was a large XSS exposure: dashboard JavaScript ran on the session's origin and every staff member could read every dashboard.
- Rebuild: one remote MCP connector replaced copied upload scripts; AI uploads land as drafts a person must publish; each dashboard runs in a sandboxed iframe on its own origin with no cookies and a strict CSP; one-hour tokens scoped to one dashboard, viewer and host; an async metadata service flags or blocks risky scripts and WebRTC.
- Totals: 325 commits, about 74,000 lines (20,000 tests), roughly $6.4k of Claude tokens over 8 months.
- Run cost: about $6k a year in AWS plus about $1.3k a month of Claude for new features, "call it $20k a year". A dashboard built by the analyst agent takes about 1 million tokens, about $0.50 after a model switch.
- Some teams still use Looker; not everything has a replacement.
How they got it
First-person account with costs from AWS bills and logged agent runs. The Looker price is an average from 355 deals reported by Vendr, not the company's actual invoice.
Limits
- Engineering time for 3 engineers over 8 months is excluded; the author says the maths changes if you must hire a team.
- Assumes a mature platform already in place (Snowflake roles, dbt, Kubernetes, Terraform).
- Usage figures are self-reported; no measure of dashboard accuracy or of data errors avoided.
- The Looker cost is an industry average, labelled as an assumption.
Why it matters for companies
The post separates what a BI licence buys: charting, permissions, semantics and hosting. Where permissions and semantics already live in the warehouse and transformation layer, AI-generated dashboards can cover much of the charting part at low cost. The security section is the part to copy: treat every generated dashboard as hostile code, isolate it, run queries as the viewer, and require a human to publish. Firms without a strong data platform team should not read the $20k figure as their cost.
Summary based on the full text. Read the original for the full argument and data.Read the original on david.lequin.dev
- Tags
- Cost