Kallisto-OSINTer
An LLM-based OSINT tool that orchestrates multiple web agents to perform deep web searches — “public intelligence done by artificial intelligence.” It is the multi-agent pole of this spoke’s AI-OSINT pattern: where llm-osint wraps a single LLM around a research loop, Kallisto distributes the work across specialized agents. By avrtt; MIT; ~8★. T3: a first-party project repo with self-reported capabilities.
What it does
Crawls the open web to aggregate public data, then applies LLMs to analyze it. Capabilities include person lookups with psychological profiling, username searches across many sites, IP/DNS analysis, and customizable visualizations.
Multi-agent architecture
- Search agent — scrapes and extracts relevant content.
- Knowledge-orchestration agent — directs the investigation flow.
- Specialized deep-dive agents — recurse into specific topics.
- LLM map-reduce — chunks and summarizes large pages token-efficiently.
- GPT-4 for the natural-language tasks.
Stack: Python, LangChain, BeautifulSoup4, PyQt6 GUI, async (aiohttp/httpx); depends on the Serper search API and ScrapingBee for bot-detection bypass.
Why it matters here
It is the orchestration step beyond single-LLM OSINT: recursive depth-seeking, token-efficient summarization, and a multi-stage prompt architecture that (its author argues) cut cost and improve investigative thoroughness versus a monolithic LLM call. Paired with llm-osint, the two mark a clear trajectory — single LLM → multi-agent orchestration — in automating reconnaissance.
Ethics note
“Person lookups with psychological profiling” is exactly the dual-use / privacy-sensitive capability the spoke flags: powerful for authorized investigation, hazardous as surveillance. Documented as part of the field, not endorsed — see synthesis.
Related
ai-osint · llm-osint · osint · social-engineering · synthesis