Open-source agentic data engineering harness providing deterministic SQL analysis, column-level lineage, dbt integration, FinOps, and multi-warehouse connections for any LLM.
altimate-code is an open-source agentic data engineering harness maintained by Altimate AI, forked from OpenCode and deeply customized for data teams. Its core design overlays a deterministic tool layer (non-LLM pattern matching) on top of any LLM to address accuracy and safety gaps of general-purpose coding agents in data engineering scenarios.
The project provides three agentic modes—Builder (read-write), Analyst (read-only), and Plan (minimal)—with hard blocks on dangerous operations like DROP DATABASE/SCHEMA/TRUNCATE. The deterministic tool layer covers SQL anti-pattern detection (19 rules, 100% F1 on benchmark), column-level lineage extraction (100% edge matching across dialects), SQL translation among 8 dialects, row-level data comparison across 12 warehouses (supporting 100M+ rows in batches), PII detection (15 categories, 30+ patterns), and FinOps cost analysis.
For dbt integration, it supports Manifest parsing, test and unit test auto-generation (dbt 1.8+, 7 dialects), model scaffolding, lineage-aware refactoring, and 12 dedicated skills. Additional capabilities include automatic chart generation from SQL results, local session tracking with HTML export, project-specific pattern training, GitHub PR/GitLab MR automated review, and one-click data stack discovery.
Architecturally, it uses a Turborepo + Bun monorepo structure, primarily in TypeScript, with MCP protocol compatibility, supporting VS Code, Claude Code, Cursor, and Windsurf. Installed globally via npm, it provides a TUI with 19 built-in skills, 20+ LLM providers, and 12 data warehouses. It ranks first on the ADE-Bench benchmark with a 74.4% score (note: benchmark methodology details not independently verified by third parties).
Installation:
npm install -g altimate-code
altimate
/connect
Unconfirmed information: README mentions "100+ deterministic tools" but only 19 skills are listed; full tool inventory requires code-level inspection. Some version dates in CHANGELOG appear anomalous and may be placeholders. No publicly referenced customer case studies found.