Altimate Code: Open-Source Agentic Data Engineering Harness

What Altimate Code Solves

General-purpose AI coding agents can write SQL but lack understanding of data context: no lineage, no schema awareness, and no comprehension of dbt manifests. According to the source, this leads to concrete problems: 27–33% of AI-generated SQL references non-existent tables, and 78% of errors are silent wrong joins that compile and run but return incorrect data. One team incurred a $5k bill from a single Cortex AI query that resource monitors missed. The issue isn't model quality but a missing tool layer.

Key Features and Capabilities

• Live column-level lineage: Traces columns through joins, CTEs, and subqueries deterministically. Achieves 100% edge match on 500K benchmark queries at 0.26ms/query without cached manifests.
• SQL anti-pattern detection: 26 rules with zero false positives at 0.48ms/query.
• Local SQL validation: Interrogates schema catalogs in 2ms without warehouse access, catching wrong tables with fuzzy-matched fix suggestions.
• Purpose-built skills: For dbt development, testing, troubleshooting, documentation, SQL optimization, and migration.
• Three agent modes: Builder, Analyst, and Planner with compiled permission enforcement. Analyst mode enforces read-only at engine level for production safety.
• Persistent memory: Cross-session with global preferences and project knowledge scopes, versioned in git and team-inherited on git pull.
• Security features: PII detection, SQL injection scanning, and permission enforcement at engine level.
• Data connectors: 10 connectors including Snowflake, BigQuery, Databricks, PostgreSQL, Redshift, DuckDB, MySQL, and SQL Server.
• Local tracer: Tracks every LLM call, tool invocation, and warehouse credit locally without external services.

Benchmark Results

On ADE-bench (the open standard from dbt Labs):

• Altimate Code (Sonnet 4.6): 74.4%
• Cortex Code (Snowflake) (Opus 4.6): 65%
• Claude Code (baseline) (Sonnet 4.6): ~40%

The source notes that a cheaper model with compiled tools outperformed a more expensive model without them, attributing the gap to the harness.

Installation and Usage

Install via npm: npm install -g @altimateai/altimate-code

Setup steps:

1. Configure LLM provider: Run altimate then /connect for interactive setup, or set environment variables like export ANTHROPIC_API_KEY=your_key.
2. Auto-detect data stack: Run altimate /discover to interrogate dbt projects, warehouse connections, and installed tools automatically.

For headless/scripted usage: altimate --yolo auto-approves all permission prompts (not recommended with live warehouse connections).

The tool integrates with existing agents via /configure-claude or /configure-codex commands.

Technical Foundation

Altimate Code is a fork of OpenCode rebuilt with a complete data layer: compiled Rust engines, purpose-built skills, and harness wiring. It's model-agnostic—works with any LLM or locally with Ollama.