Skip to main content
Your AI analyst can be confidently wrong. Nodal makes every answer accountable. A query that returns a confident, well-formed, wrong answer raises no error, no alert, and no objection on its way into a decision — we call that silent SQL. Nodal is the evaluation layer for AI analytics: see what every agent told your business, test the questions that matter against governed ground truth, and catch answer drift when data, definitions, or models change. It starts with context — what your terms mean, which table is canonical, what the standard filters are, and where the landmines are — built with your analyst, verified against the dashboards you already trust, and served to your whole team over MCP.
Prefer to watch? A video series walks through the whole flow — from why the context is interview-built to sharing the hosted MCP with your team: why interview-built context · live demo · add your team. For installation and setup, use the Quickstart.
Nodal never needs access to your database. Your agent queries your warehouse through your own read-only connector. Nodal serves the context — definitions and canonical queries — not your data. See exactly what we access →

​
Three ways to use Nodal

Nodal Context — open source, free

The Analytics Context Format (ACF) and seven agent skills — an interview that builds your context layer one domain at a time, plan/verify/challenge workflows for answering questions against it, live dashboard verification, and a local eval harness that proves the context makes the agent more accurate. Apache-2.0. Runs on your stack with Claude Code, Codex, or Cursor. Quickstart →

Nodal MCP — team

A hosted, team-shared MCP endpoint that serves the same context to everyone’s agent, so a non-technical user gets the answer the analyst would give. One .mcp.json file, OAuth sign-in, GitHub access only — never your database. Self-serve: $49/month flat for up to 20 users.

Enterprise

For teams whose decisions depend on the answers — hosted by us or deployed in your own environment — with observability of every question, regression tests on every dbt, doc, prompt, or model change, and drift detection pinned to the commit that caused it. Contact us.

​
Tested in public

Nodal’s claims are reproducible, not asserted:
  • Spider 2.0-Lite — 61.2% on 547 questions. A thin Claude Code + Sonnet harness with the right context, no fine-tuning, no custom analytics agent.
  • Shorelane — 4 of 92 answers drifted after one innocuous dbt commit. Nodal replayed every governed question and named the commit.
  • Two agents, one number. Claude Code and Codex, the same question through the same governed context, landing on the same answer.
Shorelane Commerce is our open reliability testbed: a fictional company with public BigQuery data, dbt models, dashboards, deliberately ambiguous revenue definitions, and a completed analytics-context repo anyone can inspect. Everything in these docs can be tried against it before you touch your own warehouse.

​
Why interview-built, not auto-generated

The obvious approach — ingest your warehouse, dbt, and query history and auto-generate the context — is what most tools do. Teams who have measured it found it doesn’t hold up as a source of truth:
  • Anthropic’s data team reported that auto-generating metric definitions “encoded the very ambiguities we were trying to eliminate” and was net-negative on evals versus a smaller, human-curated layer.
  • Giving an agent grep access to thousands of prior queries moved accuracy less than a point — the information was present, the agent saw it, and it still resolved questions to the wrong entity.
  • MIT CISR independently recommends the opposite of boiling the ocean: build the semantic layer incrementally, priority data assets first — which is exactly what the one-domain-at-a-time interview does.
Their conclusion, and ours: generate the draft with the model, but have a human own the definition. Nodal auto-extracts your schema, dbt, and query history as a draft to correct so the analyst isn’t staring at a blank page — but the analyst’s confirmations, not the extraction, are what we trust.
A context layer is one idea; ACF is our opinionated, eval-backed take on it. You are not required to use ACF to get value from Nodal — the eval harness reads ACF, dbt models/docs, or plain markdown. Bring whatever context you already have.

​
Evaluation is built as you go

Every disambiguation the analyst makes in the interview (“active client means X, not Y”) is simultaneously a context entry and a labeled eval pair. Each domain closes by answering real questions context-off and context-on against your live warehouse and reconciling the result with a dashboard you trust — so you can always show accuracy with the context versus without it. Watch a live verification →

​
Where to start

Quickstart

Install the plugin, connect a read-only warehouse, and run your first domain.

Share it with your team

Put the context in front of everyone’s agent with a single .mcp.json file.