lajosdeme/mole

What it solves

mole is a deep-research agent designed to solve the common pitfalls of AI-driven research: uncontrolled API costs, hallucinations (claims without evidence), and privacy risks when analyzing local data. It ensures that every claim made in a research report is backed by a verbatim quote from a source, and that the total cost of a research run never exceeds a user-defined budget.

How it works

The agent follows a structured pipeline: it decomposes a question into sub-questions, searches for sources, extracts claims, and verifies those claims against the original text. It uses a claim graph to adjudicate contradictions and synthesize a final answer with citations.

Key mechanisms include:

• Budget Enforcement: Every model call is reserved against a ledger before execution and settled after, ensuring a hard spending ceiling.
• Verification: Claims are discarded if their supporting quote does not appear verbatim in the source page.
• Privacy Boundary: For local data (CSV, JSON, etc.), the model only selects templates and column names; mole renders and runs the SQL locally, returning only aggregates (counts, means, etc.) to the model, so raw rows never leave the machine.
• MCP Integration: It can run as a standalone agent or in "toolkit mode" via the Model Context Protocol (MCP), allowing other coding agents to use its deterministic tools for research and data analysis.

Who it’s for

Researchers, developers, and analysts who need high-fidelity, verifiable AI research with strict cost controls and the ability to analyze local datasets without compromising data privacy.

Highlights

• Hard Budget Ceiling: 0% overshoot on spending limits.
• Verbatim Quote Verification: Every claim is checked against its source to eliminate hallucinations.
• Privacy-Preserving Local Analysis: Runs SQL locally on your data and only shares aggregates with the LLM.
• MCP Support: Can be driven by other agents or act as a set of research tools.
• Academic Integration: Specialized actors for querying Crossref, OpenAlex, arXiv, and PubMed.
• Dataset Mode: Ability to build structured datasets from research instead of just prose.