Robin: AI Powered Dark Web OSINT

What Robin: AI-Powered Dark Web OSINT MCP server does

Robin: AI-Powered Dark Web OSINT connects an AI-assisted investigation workflow to dark web search engines and onion pages through Tor. It refines investigation queries, searches available dark web sources, filters returned results, scrapes relevant pages, and creates an investigation summary. The resulting workflow is intended for lawful and educational OSINT use.

It supports two operating modes. A Streamlit web interface provides interactive investigations, follow-up questions grounded in the current investigation data, and suggested pivot queries. As an MCP server, Robin exposes the investigation workflow to an MCP-capable agent and uses that agent’s model.

How it works

Robin: AI-Powered Dark Web OSINT sends dark web searches through a running Tor service. The configured model helps refine queries, evaluate search results, and produce a report from the collected material. Follow-up questions can use the existing investigation data instead of starting the search again, while suggested pivots can launch a new investigation with a related query.

Model access is flexible. The MCP mode lets the connected agent provide its own model rather than requiring Robin to select one through the web interface.

Setup and configuration

A working Tor installation is required before searching. The documented installation paths include installing Tor with the system package manager on Linux or WSL, or with Homebrew on macOS, then confirming that the service is running.

Docker is the recommended deployment method. Pull the published image, mount a .env file containing a provider key, and expose port 8501 for the web UI. Robin lists models for providers whose keys it finds. Supported environment variables include OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, MISTRAL_API_KEY, and OPENROUTER_API_KEY.

Ollama does not require a key in .env, but a Docker deployment must be able to reach the Ollama service. The documented setup uses host.docker.internal and requires Ollama to listen beyond its default loopback address. Other OpenAI-compatible providers can be entered through the web UI with a base URL, optional key, and optional model name.

For MCP use, run the image with the mcp argument and connect it through the host’s MCP configuration. The README provides commands for Claude Code and Codex. Tor investigations may take several minutes, so the examples increase MCP startup and tool timeouts. Mount /app/investigations if saved reports need to persist or be shared with the web UI.

Tools and capabilities

• Search dark web engines over Tor.
• Scrape pages found during an investigation.
• Refine queries and filter search results with an LLM.
• Generate investigation summaries and save investigations.
• Ask grounded follow-up questions about collected investigation data.
• Launch fresh investigations from suggested pivot queries.
• Use hosted, local, or OpenAI-compatible model providers.
• Run through Streamlit or as an MCP service.

Limitations and notes

Robin: AI-Powered Dark Web OSINT depends on Tor for searches, so availability and response time depend on the Tor connection and the onion sources it reaches.

The project warns that dark web access and interaction may be illegal in some jurisdictions. Use it only for lawful, authorized investigations and follow applicable institutional policies. Queries may be sent to third-party LLM or API providers, so sensitive information should be reviewed before submission and provider terms should be checked.

Docker persistence also has filesystem considerations. A mounted investigations directory must be writable by the container’s UID, or saved reports can fail even though the investigation itself runs. A named Docker volume is provided as an alternative.