Contextloop
Serves your team's metric definitions to data agents, turns user corrections into approved rules, and tracks SQL answer accuracy weekly.
Hosted MCP Server
npx add-mcp 'https://getcontextloop.com/mcp'Installs into Claude Code, Codex, Cursor and more
Documentation
Your data agent gets the metrics wrong.Contextloop makes it learn.
Claude and Cursor can write SQL against your warehouse, but they don't know that "revenue" excludes refunds or that "active" means 2+ sessions in 28 days. Contextloop serves your team's definitions to every agent over MCP, turns each correction into a rule, and tracks how often answers are right.
Try it now without an account: claude mcp add --transport http contextloop-demo https://getcontextloop.com/mcp?key=demo
Works with Claude Code, Claude Desktop, Cursor, and any MCP client. We never need access to your warehouse.
You: What was net revenue last month?
Agent get_context: "Net revenue" = sum(amount) where status not in ('refunded','test'), in USD using fx_rates. Last month: $412,880.
You: Close, but we also exclude internal accounts (accounts.is_internal).
Agent report_outcome: Correction saved. Corrected answer: $398,210. Next time, anyone on your team asking gets the right number on the first try.
How it works
1. Connect your agent
Paste one MCP URL into Claude Code, Cursor or Claude Desktop. Your agent keeps using its own database tools. Contextloop only supplies context, so no warehouse credentials are needed.
2. Seed or import definitions
Add your key metrics, terms, table notes and rules, or paste in existing dbt / Cube YAML or a glossary doc. Agents call get_context before writing SQL.
3. Corrections become rules
When someone says "no, that should exclude trials", the agent files the correction. You approve it in one click and every agent on the team picks it up.
4. Watch accuracy climb
Every question is logged as correct or corrected. A weekly accuracy chart shows whether your agents are actually getting better. Export to dbt MetricFlow or Cube YAML at any time, so you're never locked in.
Why not just a YAML file in the repo?
It goes stale
Hand-written glossaries drift. Contextloop is fed by the questions your team actually asks and the corrections they actually make.
It's per-person
A CLAUDE.md on one laptop doesn't help the analyst next to you. One workspace serves every agent on the team.
Nothing measures it
Without a scorecard you can't tell whether the agent is right 60% or 95% of the time. Contextloop tracks it per week.
Pricing
Free
$0
1 workspace, up to 50 definitions, unlimited agent calls. Enough to see if accuracy moves.
Team
$29/month
Unlimited definitions, shared workspace for your whole data team, approval workflow, weekly accuracy report by email, YAML export.
Start with Free, upgrade later
Accuracy pilot
Free
Send us ~50 real questions your team asked an agent. We'll build the definitions and show before/after accuracy on your own questions.