OpenOSINT is an open source intelligence (OSINT) agent for security researchers and analysts. It puts 21 investigation tools behind a natural-language interface: you describe a target (an email address, username, domain, IP address, name, or phone number) and the agent selects tools, runs them, and chains further lookups on what it finds, such as pivoting from a breach record to a GitHub username. Every finding is shown with the tool call that produced it, and AI summaries are presented as leads to verify rather than proof.
It runs as an interactive REPL, a CLI for direct tool calls without AI, a browser Web UI with streaming chat and result cards, and an MCP (Model Context Protocol) server that any MCP-compatible client such as Claude Code, Claude Desktop, or Cursor can drive. The AI backend is Anthropic Claude by default, with local inference through Ollama or any OpenAI-compatible endpoint (LiteLLM, vLLM, LM Studio) as alternatives. Three of the tools (email, username, and domain recon) are also hosted as pay-per-use Apify Actors, reachable from a browser or as MCP tools through the Apify MCP Server with no server to run.
The agent starts in passive mode: it and any MCP client are offered only tools that query third-party data sources or compute locally. Tools that touch the target's infrastructure or send many requests from your IP (holehe email probes, sherlock username checks, sublist3r subdomain scraping, phoneinfoga, URL scraping, and VirusTotal URL submission) are not listed to the model until you opt in, each tool description carries its noise level, and a publicly reachable instance can never enable them. A per-request tool-call cap stops an investigation that keeps calling tools.
| Tool | Source | Investigates |
|---|---|---|
| search_email | holehe | accounts linked to an email address |
| search_username | sherlock | username presence across 400+ platforms |
| search_breach | HaveIBeenPwned | data breach exposure |
| search_whois, search_rdap | WHOIS, RDAP | registrar, dates, name servers |
| search_ip, search_ip2location | ipinfo.io, IP2Location.io | geolocation, ASN, VPN/proxy/Tor flags |
| search_domain | sublist3r | subdomain enumeration |
| search_dns | dnspython | DNS records, SPF, DMARC, DKIM |
| search_shodan, search_censys | Shodan, Censys | open ports, banners, CVEs, certificates |
| search_virustotal, search_abuseipdb | VirusTotal, AbuseIPDB | malware verdicts, IP abuse reputation |
| search_github | GitHub REST API | profile, repos, commit-discovered emails |
| search_paste, search_phone | psbdmp.ws, phoneinfoga | paste dumps, carrier and line type |
| generate_dorks, search_dorks_live, search_footprint | built-in, Bright Data SERP | dork URLs, live results, search-engine footprint |
| scrape_url | Bright Data Web Unlocker | page content as Markdown |
| search_gdelt_geo | GDELT feed | recent news placed on a map |
An optional graph module adds a FollowTheMoney entity graph with statement-level provenance, an append-only store, non-destructive same_as deduplication candidates for human review, and .ftm export. Stated limits: nine of the 21 tools require a third-party API key and inherit that provider's rate limits and pricing, the local binaries run as unsandboxed subprocesses, and a negative result means the source reported no match, not that no exposure exists.
Features
- AI tool chaining: the agent picks tools from a plain-language description of the target and chains further lookups on its findings
- 21 tools: email, username, breach, WHOIS, RDAP, IP, subdomain, dorks, paste sites, phone, Shodan, VirusTotal, Censys, IP2Location, AbuseIPDB, GitHub, DNS with SPF, DMARC, and DKIM checks, GDELT news geolocation, live dork search, URL scraping, and search-engine footprint
- Three AI backends: Anthropic Claude by default, local Ollama, or any OpenAI-compatible endpoint such as LiteLLM, vLLM, or LM Studio
- MCP server: 20 passive tools listed by default and 25 with active tools enabled, for Claude Code, Claude Desktop, Cursor, and other MCP clients
- Passive by default: tools that touch the target or are noisy stay hidden from the model until you opt in, each labeled with its noise level, and a public instance can never enable them
- Parallel execution: complementary tools run concurrently within an investigation
- Reports and history: PDF and Markdown reports are saved after every investigation, and every REPL session is saved and browsable
- Web UI: browser chat with streaming tool output, result cards, a light and dark theme, and a client-side entity correlation graph of the current chat
- Entity graph module: an optional FollowTheMoney store with statement-level provenance, a same_as review queue, and .ftm export
- Hosted Actors: email, username, and domain recon run as pay-per-use Apify Actors, usable as MCP tools without running a server
- Multi-target investigation: investigate_multi runs one agent per target under a shared tool-call cap
