Smart engineers shouldn't spend their evenings chasing breakage across vendor walls. Parrat puts the toil where it belongs: on the agent, not the human.

Parrat is an open-source CLI that uses Claude to investigate data incidents. Point it at your data stack, describe the problem, and get a root cause in under 90 seconds — with a full audit trail.

Validated across 11 live investigations with 100% correct root causes at an average cost of $0.07 per investigation.

How it works

Parrat runs Skills — pre-codified investigation playbooks that reason across your stack using a deliberately thin set of tools. Each Skill gives Claude access to only the tools it needs for that specific investigation, producing predictable, auditable reasoning paths.

Every run writes to an append-only audit log. Every run is replayable.

parrat run freshness-investigation
parrat run metric-drop-rca
parrat run lineage-analysis

Skills

| Skill | What it investigates | |---|---| | freshness-investigation | Why is this source stale? Which downstream models are at risk? | | metric-drop-rca | Why did this metric drop? Which upstream model caused it? | | lineage-analysis | What does this model depend on, and what depends on it? |

Prerequisites

• Node.js 20+
• dbt-mcp running and accessible
• ANTHROPIC_API_KEY (set in environment or .env file)

Install

npm install -g @parrat/cli

Configure

Create a parrat.config.yaml in your project root:

mcp:
  command: uvx
  args: ["dbt-mcp"]

skills:
  freshness-investigation:
    enabled: true
  metric-drop-rca:
    enabled: true
  lineage-analysis:
    enabled: true

See examples/snowflake/ for a complete configuration example.

Verify your setup

parrat doctor

This checks Node version, API key, config file, and dbt-mcp connectivity before you run your first investigation.

Run an investigation

parrat run freshness-investigation

Parrat reads your data stack, reasons about what broke, and returns a structured root cause with confidence rating. The full reasoning chain is logged to .parrat/audit.jsonl.

Replay any investigation

parrat replay <run_id>

Every investigation is replayable — every tool call, every Claude turn, input tokens, output tokens, cost, and duration.

Examples

• examples/snowflake/ — configuration for Snowflake + dbt
• examples/custom-skill/ — write your own investigation Skill in TypeScript

License

Apache 2.0 — see LICENSE.

Built by Raguvind Tharanitharan.