RELATED 01
ketch
Web search, public code search, library docs and page-to-markdown scraping from one Go binary
742GitHub stars
1.1k/wknpm downloads
48popularity score
Robin
Dark web OSINT over Tor: query onion search engines, read .onion pages and save reports
By Apurv Singh Gautam (individual) Python Local · stdio SecurityWeb searchBrowser automation
69 popularity score ★ 7.9k ⇣ 68k pulls
Connect your client
Local (Docker)
Add in one click
Opens the app, which asks before it adds the server.
Or add it by command or config
Claude CodeTerminal
claude mcp add robin -- docker run -i --rm -v robin-investigations:/app/investigations apurvsg/robin mcpClaude Desktopclaude_desktop_config.json
{
"mcpServers": {
"robin": {
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"-v",
"robin-investigations:/app/investigations",
"apurvsg/robin",
"mcp"
]
}
}
}Cursor~/.cursor/mcp.json
{
"mcpServers": {
"robin": {
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"-v",
"robin-investigations:/app/investigations",
"apurvsg/robin",
"mcp"
]
}
}
}VS Code.vscode/mcp.json
{
"servers": {
"robin": {
"type": "stdio",
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"-v",
"robin-investigations:/app/investigations",
"apurvsg/robin",
"mcp"
]
}
}
}Codex~/.codex/config.toml
[mcp_servers.robin]
command = "docker"
args = ["run", "-i", "--rm", "-v", "robin-investigations:/app/investigations", "apurvsg/robin", "mcp"]Gemini CLITerminal
gemini mcp add --scope user robin docker -- run -i --rm -v robin-investigations:/app/investigations apurvsg/robin mcpSettings 6
ROBIN_MODEL |
Optional: Model robin_investigate uses when the client offers no sampling. |
|---|---|
OPENAI_API_KEY |
Optional: OpenAI key; ANTHROPIC_API_KEY, GOOGLE_API_KEY, MISTRAL_API_KEY or OPENROUTER_API_KEY work instead. |
OLLAMA_BASE_URL |
Optional: Ollama address; defaults to http://host.docker.internal:11434. |
ROBIN_DEFAULT_PRESET |
Optional: Default research domain; threat_intel unless set. |
ROBIN_DEFAULT_MAX_RESULTS |
Optional: Default result count, 10 to 100 (50 if unset). |
ROBIN_DEFAULT_MAX_SCRAPE |
Optional: Default pages to read, 3 to 20 (10 if unset). |
In a container, a setting goes in with -e NAME=value after docker run.
What it does
Robin is a dark web research tool that ships as one Docker image with two modes: a Streamlit web UI on port 8501, and an MCP server on stdio started with the `mcp` argument. The container starts its own Tor daemon, and every search and page fetch goes out through it, so the MCP client never loads an onion page itself.
In MCP mode the assistant works through the steps of an investigation: turn a question into a short search query, send it to sixteen onion search engines, drop off-topic hits, fetch the remaining pages as text, then write a report and save it as JSON in a Docker volume the web UI can also open. Scraped text comes back wrapped in marked untrusted-data delimiters, with control and invisible characters stripped first. The refine, filter and investigate steps run on the client's own model through MCP sampling when the client supports it. Otherwise they need ROBIN_MODEL and a provider key. Without either, the step-by-step tools still work, because the assistant writes the report itself.
The server has nine tools in v3.1 (listed in TOOL_NAMES in mcp_server.py). A search does not run until the user has picked a research domain (preset) and the call passes `user_chose=true`. Saving is rate-limited to 30 saves per 10 minutes, and a saved report is capped at 200,000 characters, 100 sources and 10 pivots (mcp_server.py, v3.1).
What it needs
| Runtime | Docker |
|---|
Docker, with the image apurvsg/robin pulled first so the first start is not a download. Tor runs inside the container. No key is needed when the MCP client supports sampling or you drive the tools step by step. For robin_investigate without sampling, pass ROBIN_MODEL plus one provider key (OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY, MISTRAL_API_KEY or OPENROUTER_API_KEY) into the container, or point it at a local Ollama.
Limits
Its makers say it is for educational and lawful investigation only, and that opening some dark web content may be illegal where you live. A Tor investigation takes minutes, so the README raises client tool timeouts to 600 seconds. Until Tor has finished bootstrapping, search tools return a Tor-not-ready message. robin_scrape rejects hosts that are not onion services unless `allow_clearweb` is set. robin_investigate defaults to 50 results, 10 pages and 8,000 characters per page (config.py). For Claude Code, scrape output is capped at 60,000 characters unless `max_chars` is given.
Who it suits
Suits threat-intelligence and OSINT analysts who want an assistant to search onion sites and gather dark web evidence without opening Tor themselves. It works with clients that can run a local Docker command. The README says an ordinary ChatGPT chat cannot run it, only Codex sessions. Not suited to clearweb research or to people who cannot run Docker. Teams that must not send scraped dark web text to an outside model provider should use the client's own model or a local Ollama.
Tools 9
robin_refine- Turns a question into a short query suited to onion search engines
robin_search- Queries sixteen onion search engines over Tor and returns results with short ids
robin_filter- Keeps only the results of the latest search that are on the query's topic
robin_scrape- Fetches filtered results or given .onion URLs over Tor and returns their text
robin_investigate- Runs refine, search, filter, scrape, summary and save in one call
robin_save_investigation- Saves a report the assistant wrote, with sources and pivots, to the investigations volume
robin_list_investigations- Lists the investigations saved so far
robin_list_models- Lists the models Robin can use with the keys it has
robin_health- Reports whether Tor is up and can probe a model or the search engines