agentic-analytics
Open-source, self-hosted AI analyst for business files. Give it spreadsheets, CSV exports, PDFs, Word documents and emails, and it reports what changed, why, and what to do next. Every number is computed by SQL and rechecked by a verifier before the report ships.
- Category
- AI & Automation
- Year
- 2026
Inside agentic-analytics
Chat AIs often do the math in their heads and state the result with confidence. agentic-analytics never lets the model do the math: numbers come from SQL in DuckDB, a separate verifier recomputes them, and every claim links to its source row, cell, page or URL. It runs locally for free with Ollama or a free endpoint, and plugs in through a CLI, an HTTP API, an MCP server or Telegram. Apache-2.0, v0.1, early.
The problem
Chat AIs often do the math in their heads and state the result with confidence. agentic-analytics is an open-source project built around one rule: the model never does the math.
What I built
A self-hosted AI analyst for business files. Give it spreadsheets, CSV exports, PDFs, Word documents and emails, and it reports what changed, why, and what to do next. Every claim links to its source row, cell, page or URL.
How it works
Numbers come from SQL in DuckDB, never from the model, and a separate verifier recomputes them before the report ships. It runs locally for free with Ollama or a free model endpoint, and plugs in through a CLI, an HTTP API, an MCP server or Telegram.
Stack
- TypeScriptFrontend
- DuckDBDatabase
- HonoBackend
- Vercel AI SDKAI / ML
- OllamaAI / ML
- MCPAI / ML
Status
- State
- Open source
- Year
- 2026
Results
Open source under Apache-2.0, at version 0.1 and early. On a 30-job eval suite it produced 0 number mismatches.
Key facts
- Numbers from SQL, never from the model
- Every claim cited to its source
- 0 number mismatches on a 30-job eval suite
- Runs locally, free models supported
- CLI, API, MCP and Telegram
- Apache-2.0 licensed