Author here, Kyle, founder of ClariLayer (ex-Databricks/Cloudflare).
I was excited when I saw that both Anthropic and OpenAI shared the same view about the importance of context for data agents. No doubt that better context could beat a smarter model in a lot of data-agent use cases. These two well-resourced companies show us what the destination of context-layer building could look like, but the reality is that most organizations out there don't have the resources and a full data team to build out the context layer they described, let alone the ongoing maintenance.
I built ClariLayer to solve exactly this. It is a lightweight MCP server that any individual analyst can connect to their Claude Code/Codex/Cursor. So we all can have a portable context layer across tools, projects, and sessions. And it's completely free for all individual analysts. No card needed, npx clarilayer init.
Two things it deliberately does not do. There is no "verified" status: we built that stamp, then gated it off before launch because we couldn't make it sound against the long tail of SQL dialects (the last section of the essay is about this). Saved context is only ever asserted or carries a caveat. And it never holds your warehouse credentials or runs SQL on our side; your agent runs the checks with its own access.
Happy to answer anything about the failure stories in the essay. And if you try it on your warehouse, I'd like to hear where it breaks.
Author here, Kyle, founder of ClariLayer (ex-Databricks/Cloudflare).
I was excited when I saw that both Anthropic and OpenAI shared the same view about the importance of context for data agents. No doubt that better context could beat a smarter model in a lot of data-agent use cases. These two well-resourced companies show us what the destination of context-layer building could look like, but the reality is that most organizations out there don't have the resources and a full data team to build out the context layer they described, let alone the ongoing maintenance.
I built ClariLayer to solve exactly this. It is a lightweight MCP server that any individual analyst can connect to their Claude Code/Codex/Cursor. So we all can have a portable context layer across tools, projects, and sessions. And it's completely free for all individual analysts. No card needed, npx clarilayer init.
Two things it deliberately does not do. There is no "verified" status: we built that stamp, then gated it off before launch because we couldn't make it sound against the long tail of SQL dialects (the last section of the essay is about this). Saved context is only ever asserted or carries a caveat. And it never holds your warehouse credentials or runs SQL on our side; your agent runs the checks with its own access.
Happy to answer anything about the failure stories in the essay. And if you try it on your warehouse, I'd like to hear where it breaks.