Eidolon

Eidolon is a privacy-first approach to finding and understanding your digital footprint. The stack uses OSINT tools to gather information based on your search parameters, then a local LLM compiles it into a report. The LLM runs on your machine, so no data ever reaches an external service or leaves your box — you own your data.

What it does

• Aggregates ~25 OSINT sources — breaches, leaked credentials, data brokers, account enumeration, public records, phone/email intel, exposed hosts — into a single risk report.
• Maps findings to MITRE ATT&CK so you see what an attacker could actually do with what's exposed.
• Flags AI-training exposure (which platforms may train on your data, and how to opt out).
• Risk scoring and the leaked-credential dossier are deterministic — built from scan state, not the LLM — so the report survives an LLM hiccup. The model only writes narrative.
• Outputs Markdown, PDF, and JSON.

MCP-native

Eidolon runs as an MCP server, so you can drive it from any MCP client (Claude Desktop, Claude Code) — scan a target, list past scans, and read reports conversationally. It runs locally over stdio; your data never leaves the box.

Tools: scan_target, scan_status, list_scans, get_report, reveal_credentials. A scan takes minutes, so scan_target returns a scan_id immediately and runs in the background — poll scan_status(scan_id) until it reports done, then get_report(scan_id). The leaked-credential dossier (plaintext passwords from breach dumps) is redacted by default and only returned when you explicitly call reveal_credentials.

Requirements

Everything below is optional — Eidolon runs with whatever you give it and tells you, per source, what it couldn't check (no token) versus what it checked and found nothing.

• Python 3.11+ and uv (required)
• Ollama for the local LLM narrative: ollama pull llama3.1:8b — if it's down, the report is still produced deterministically (just no written narrative).
• A running SpiderFoot instance — skipped if unreachable.
• API keys — see .env.example. Each unlocks one data source and skips cleanly if absent (the report says "not checked — set X"). A scan with no keys is sparse but still runs.

New here? Start with the Setup & Configuration guide — which keys to get, what each one costs, and what it unlocks.

Quickstart

git clone https://github.com/sudohnim/eidolon && cd eidolon
uv sync
cp .env.example .env        # fill in your keys

# scan yourself from the CLI
uv run eidolon --email you@example.com

# or run the MCP server (stdio)
uv run eidolon-mcp

Use from Claude Desktop

Option 1 — from a local clone (loads .env from cwd):

{
  "mcpServers": {
    "eidolon": {
      "command": "uv",
      "args": ["run", "eidolon-mcp"],
      "cwd": "/absolute/path/to/eidolon"
    }
  }
}

Option 2 — from PyPI via uvx (recommended, no clone needed). Pass your keys in the env block:

{
  "mcpServers": {
    "eidolon": {
      "command": "uvx",
      "args": ["--from", "eidolon-osint", "eidolon-mcp"],
      "env": {
        "HIBP_API_KEY": "...",
        "DEHASHED_API_KEY": "...",
        "OLLAMA_HOST": "http://localhost:11434"
      }
    }
  }
}

Option 3 — straight from GitHub (pre-release / latest commit):

{
  "mcpServers": {
    "eidolon": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/sudohnim/eidolon", "eidolon-mcp"],
      "env": {
        "HIBP_API_KEY": "...",
        "DEHASHED_API_KEY": "...",
        "OLLAMA_HOST": "http://localhost:11434"
      }
    }
  }
}

Add only the keys you have — the rest skip cleanly. Then ask Claude to "scan my email"; it calls scan_target, polls scan_status, and reads the report.

How it works

A LangGraph pipeline: intake → wave 1 scans → wave 2 scans → MITRE mapping → correlation → analysis → report. Both the CLI and the MCP server call the same run_scan() core; reads go through a small repository layer. See docs/ROADMAP.md for the architecture and where it's headed (stateful history, continuous monitoring).

Intended use

Eidolon is for scanning yourself, or targets you are explicitly authorized to assess (authorized security testing, your own footprint). It surfaces real secrets, including plaintext passwords from breach dumps. Do not use it to profile or surveil people without their consent. You are responsible for complying with the terms of the data sources you configure and with applicable law.

License

AGPL-3.0. If you run a modified version as a network service, you must offer users its source.