robin

Architecture

Related MCP server: darknet-mcp-server

Features

• 🤖 Multi-Model Support – OpenAI, Claude, Gemini, Mistral, OpenRouter, Ollama, or any OpenAI-compatible API (LM Studio, llama.cpp, Groq, etc.).

• 🌐 Web UI – Streamlit-based interface for interactive investigations.

• 🔌 MCP Support – Run Robin from any MCP-capable agent, using that agent's own model. See Robin as MCP.

• 💬 Conversational Follow-ups – Ask grounded follow-up questions about an investigation without re-running the search, answered from that investigation's own data.

• 🔀 One-Click Pivots – Suggested follow-up queries surfaced from the findings; click one to launch a fresh investigation.

• 🐳 Docker-Ready – Recommended Docker deployment for clean, isolated usage.

⚠️ Disclaimer

This tool is intended for educational and lawful investigative purposes only. Accessing or interacting with certain dark web content may be illegal depending on your jurisdiction. The author is not responsible for any misuse of this tool or the data gathered using it.

Use responsibly and at your own risk. Ensure you comply with all relevant laws and institutional policies before conducting OSINT investigations.

Additionally, Robin leverages third-party APIs (including LLMs). Be cautious when sending potentially sensitive queries, and review the terms of service for any API or model provider you use.

Installation

Docker [Recommended]

• Pull the latest Robin docker image

docker pull apurvsg/robin:latest

• Create a .env file in the folder you run from, with your API key in it (see .env.example). Create it before the first run: if it does not exist, Docker mounts an empty folder in its place and Robin starts with no keys.

touch .env   # then add your API key to it

• Run the docker image as:

docker run --rm \
   -v "$(pwd)/.env:/app/.env" \
   --add-host=host.docker.internal:host-gateway \
   -p 8501:8501 \
   apurvsg/robin:latest

docker run --rm \
   -v "$(pwd)/.env:/app/.env" \
   -v robin-investigations:/app/investigations \
   --add-host=host.docker.internal:host-gateway \
   -p 8501:8501 \
   apurvsg/robin:latest

docker run --rm \
   -v "$(pwd)/.env:/app/.env" \
   -v "$(pwd)/investigations:/app/investigations" \
   --add-host=host.docker.internal:host-gateway \
   -p 8501:8501 \
   apurvsg/robin:latest

• Open your browser and navigate to http://localhost:8501

Build it yourself

To run your own build instead of the published image, clone the repository and build it:

docker build -t robin .

Then use the same run commands with robin in place of apurvsg/robin:latest:

docker run --rm \
   -v "$(pwd)/.env:/app/.env" \
   --add-host=host.docker.internal:host-gateway \
   -p 8501:8501 \
   robin

docker run --rm \
   -v "$(pwd)/.env:/app/.env" \
   -v "$(pwd)/investigations:/app/investigations" \
   --add-host=host.docker.internal:host-gateway \
   -p 8501:8501 \
   robin

• Open your browser and navigate to http://localhost:8501

Robin as MCP

Any LLM service that speaks MCP can run it, using its own model. The sections below are worked examples for the common hosts; a host that is not listed works the same way, with whatever wording it uses for "add an MCP server".

Claude Code

claude mcp add robin -- docker run -i --rm -v robin-investigations:/app/investigations apurvsg/robin mcp

Or check it into the project in .mcp.json. A Tor investigation takes minutes, so raise the per-server timeout:

{
  "mcpServers": {
    "robin": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-v", "robin-investigations:/app/investigations", "apurvsg/robin", "mcp"],
      "timeout": 600000
    }
  }
}

Codex

codex mcp add robin -- docker run -i --rm -v robin-investigations:/app/investigations apurvsg/robin mcp

The ChatGPT desktop app shares this host: adding Robin under Settings → MCP servers there writes the same configuration, and its tools appear in Codex sessions rather than in an ordinary ChatGPT chat.

A Tor search usually takes a couple of minutes, which can outlast Codex's default MCP tool timeout, so raise the limits in ~/.codex/config.toml:

[mcp_servers.robin]
command = "docker"
args = ["run", "-i", "--rm", "-v", "robin-investigations:/app/investigations", "apurvsg/robin", "mcp"]
tool_timeout_sec = 600
startup_timeout_sec = 60

# Only for non-interactive runs (`codex exec`): Codex declines MCP tool calls
# that would need approval, so pre-approve Robin's read-only tools.
default_tools_approval_mode = "approve"

Pull the image first (docker pull apurvsg/robin:latest) so the first start is not also a download.

Claude Desktop

Add this to claude_desktop_config.json and restart the app:

{
  "mcpServers": {
    "robin": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-v", "robin-investigations:/app/investigations", "apurvsg/robin", "mcp"]
    }
  }
}

Hermes

In ~/.hermes/config.yaml:

mcp_servers:
  robin:
    command: docker
    args: [run, -i, --rm, -v, "robin-investigations:/app/investigations", apurvsg/robin, mcp]
    timeout: 600

OpenClaw

openclaw mcp add robin --command docker --arg run --arg -i --arg --rm --arg -v --arg robin-investigations:/app/investigations --arg apurvsg/robin --arg mcp

Or in the config, under mcp.servers:

{
  "mcp": {
    "servers": {
      "robin": {
        "transport": "stdio",
        "command": "docker",
        "args": ["run", "-i", "--rm", "-v", "robin-investigations:/app/investigations", "apurvsg/robin", "mcp"],
        "requestTimeoutMs": 600000
      }
    }
  }
}

ChatGPT

The desktop app's Settings → MCP servers configures the Codex host that the app, the Codex CLI and the IDE extension share, so add Robin there and it appears in Codex sessions: see Codex. An ordinary ChatGPT chat cannot run it, because chat reaches MCP servers through connectors that run in OpenAI's infrastructure rather than on your machine.

Security

• Nothing sends, executes, or runs a shell. The only writes are saved reports, into investigations/ under a filename Robin picks; the agent never names a path.

• Searches, page scrapes and search-engine health checks go through Tor, so the app you are using never fetches a dark web page itself. Model provider calls and the model list refresh go directly, not through Tor.

• Scraped text comes back inside untrusted-data delimiters, with control, zero-width and bidi characters stripped first, so your model reads it as evidence rather than as instructions.

Acknowledgements

• Idea inspiration from Thomas Roccia and his demo of Perplexity of the Dark Web.

• Tools inspiration from my OSINT Tools for the Dark Web repository.

• LLM Prompt inspiration from OSINT-Assistant repository.

• Logo Design by my friend Tanishq Rupaal