Built by an operator tired of re-explaining the data.

Analysts have moved into AI coding agents — Claude Code, Cursor, Codex — and they are productive there. But the agent has no durable memory of your data. Every session starts from scratch: which table is the real one, how two tables actually join, why this status code means a refund, that one customer who breaks every rule. You explain it, you get an answer, and tomorrow you explain it again.

The common workaround — a hand-written CLAUDE.md of definitions — helps, but it has the same trust problem as the original numbers: it is just asserted text. Nothing checks that what you wrote down still matches what the warehouse actually returns. So the agent confidently repeats a definition that quietly drifted, and you only find out when two reports disagree.

That disagreement is the acute moment, and it is the research-validated #1 analyst pain: why don't these two numbers match? Today the honest answer is a slow manual audit. ClariLayer was built to make that moment fast — to reconcile a saved definition against your real warehouse result and flag a caveat the instant the declared logic and the actual data disagree.

ClariLayer is a personal, self-deployed context layer for your AI coding agent, delivered over MCP. You install it into Claude Code, Cursor, or Codex with one command, and from then on your agent can recall the right context about your data in-flow, mid-task, without you leaving your editor. There is no destination app to visit and no team account to provision — single-player by design: one analyst, your own data, your own agent.

It runs on four verbs. Bootstrap populates your context from the SQL, dbt models, and CLAUDE.md you already have. Remember saves a new definition, join path, or gotcha so it survives across sessions. Recall pulls the most relevant context for the task at hand. And reconcile grounds a saved definition against your real warehouse result — your agent runs the SQL with its own access and reports back, and ClariLayer flags a caveat when the declared logic and the actual data disagree. We never hold your warehouse credentials and never execute SQL ourselves.

We are deliberate about what we claim. Today reconcile records a caveat on a mismatch or leaves an entry asserted — it does not stamp a definition “verified.” A single false “verified” is the one failure a trust product cannot ship, so the stronger trust mark is gated off until a real SQL parser can back it. That is the documented trajectory; the present-tense promise is reconciled, evidence-backed, caveat-aware context — checked against your source, not blindly asserted.

This is where we start, not where it ends. The same engine that grounds one analyst's context is the bridge to the team: the personal context analysts build is what later merges into shared, owned, governed team context — ownership, approval, the one right metric. That team and governance story (the original Metric Lifecycle Management thesis) is still here, kept as the “for teams” Expansion path — it is just no longer the headline.