Stealth Web Search gives any MCP client a real, JavaScript-rendering browser with built-in anti-fingerprinting, plus sub-agents that browse, automate and cross-check the web on their own. It runs in one Docker container, logs every step, and lets you watch it all live.
Quick start
You need Docker. The first build downloads the checksum-verified Obscura browser and compiles the server; it takes about a minute.
http://127.0.0.1:8931/mcp (Streamable HTTP)http://127.0.0.1:8931/127.0.0.1 only, with optional bearer-token auth./logs, one transcript per sub-agent run$ git clone https://github.com/tanishmeh/stealth-web-search.git $ cd stealth-web-search $ docker compose up -d --build $ curl -s http://127.0.0.1:8931/healthz {"ok":true,"name":"stealth-web-search","version":"0.1.0",...}
Features
A production-quality MCP server around a fast, stealthy browser engine, with the observability to see exactly what your agents did.
41 browser_* tools: navigate, read pages as text, Markdown, links or structured data, click, type, fill forms, use tabs and cookies, run JavaScript, take screenshots and PDFs. Pages run their scripts in V8 first, so single-page apps work.
Obscura's stealth build presents a consistent Chrome fingerprint, impersonates its TLS fingerprint, reports navigator.webdriver as false and blocks trackers. Element references stay on the server, so nothing is written into the page.
Hand over a task and get back only the result. Each run gets its own isolated browser and uses any OpenAI-compatible model: vLLM, LM Studio, Ollama, llama.cpp or a hosted API.
The dashboard streams the agent's tab with markers where it clicks and types, lists every tool call with its arguments and result, and shows the console, network, logs and each sub-agent's reasoning as it streams.
Every MCP message, tool call, browser command, console message and network request goes to rotating JSON log files. Passwords, cookie values and credential headers are redacted, and each sub-agent run leaves a transcript.
One click adds it to LM Studio. Claude Code, Cursor, VS Code and other clients connect over Streamable HTTP; stdio-only clients use the bundled bridge.
The automation agent turns a task into a verified script with named parameters. Replay it later with script_run in a fresh browser, with no model, in a QuickJS sandbox with memory and time limits.
Non-root, read-only file system, every Linux capability dropped. The browser's control port never leaves the container, private networks are blocked by default, and sub-agents use a separate browser engine.
Unit and integration suites run against the real browser engine, locally and in the container. Sub-agents are tested with a scripted model and live with a real one, and LM Studio scenarios run against a local model.
Sub-agents
Sub-agents run inside the container next to the browser. Your agent stays in charge and keeps its context small; the sub-agent does the clicking and reading, and returns only what you asked for.
Completes a task in its own browser and returns the output in the shape you describe, as text or validated JSON.
Does the task once, writes a script that repeats it for any inputs, verifies it in a fresh browser, and stores it.
Searches the web, reads the pages, cross-checks several independent sources and returns the answer with cited, verified quotes.
See every step
The dashboard at http://127.0.0.1:8931/ shows what each agent sees and does, while it does it.
./logs as JSON lines.
mcp/stealth-web-search integration.Integrations
Use the tools from an LM Studio chat, run the sub-agents on an LM Studio model, or connect any MCP client.
LM Studio connects to MCP servers listed in its mcp.json. Add this one in one click, or with the setup script.
docker compose up -dnpm run lmstudio:setupmcp/stealth-web-search under Integrations{
"mcpServers": {
"stealth-web-search": {
"url": "http://127.0.0.1:8931/mcp",
"timeout": 180000
}
}
}The sub-agents can run on LM Studio too: see the LM Studio provider in config/models.example.json.
The endpoint is http://127.0.0.1:8931/mcp (Streamable HTTP). Clients that only speak stdio use the bridge that ships in the image.
claude mcp add --transport http stealth-web-search http://127.0.0.1:8931/mcp{
"mcpServers": {
"stealth-web-search": {
"url": "http://127.0.0.1:8931/mcp"
}
}
}In ~/.cursor/mcp.json or .cursor/mcp.json.
{
"servers": {
"stealth-web-search": {
"type": "http",
"url": "http://127.0.0.1:8931/mcp"
}
}
}In .vscode/mcp.json in your workspace.
{
"mcpServers": {
"stealth-web-search": {
"command": "docker",
"args": ["exec", "-i", "stealth-web-search", "node", "dist/stdio-bridge.js"]
}
}
}For Claude Desktop and other stdio-only clients. The container must be running.
AUTH_TOKENModel configuration
Copy config/models.example.json to config/models.json, set your endpoint, model and key, and restart. The format is the provider list editors use for custom endpoints, so you can reuse the one you already have.
docker compose run --rm --no-deps stealth-web-search node dist/check-config.js --ping shows the settings the agents will use and whether the model answers.[
{
"name": "Local vLLM",
"vendor": "customendpoint",
"apiKey": "your-api-key",
"apiType": "chat-completions",
"models": [
{
"id": "qwen3.8-27b",
"name": "Qwen3.8 27B (vLLM)",
"url": "http://192.168.1.50:8000/v1/chat/completions",
"toolCalling": true,
"vision": false,
"streaming": true,
"contextWindow": 262144,
"maxOutputTokens": 32768,
"thinking": true,
"supportsReasoningEffort": ["low", "medium", "xhigh"],
"reasoningEffortFormat": "chat-completions",
"modelOptions": {
"temperature": 0.4,
"top_p": 0.95
}
}
],
"settings": {
"qwen3.8-27b": {
"reasoningEffort": "medium"
}
}
},
{
"name": "LM Studio",
"vendor": "lmstudio",
"apiKey": "lm-studio",
"apiType": "chat-completions",
"models": [
{
"id": "qwen/qwen3.8-27b",
"name": "Qwen3.8 27B (LM Studio)",
"url": "http://host.docker.internal:1234/v1/chat/completions",
"toolCalling": true,
"vision": true,
"streaming": true,
"contextWindow": 32768,
"maxOutputTokens": 8192,
"thinking": true,
"supportsReasoningEffort": ["low", "medium", "high"],
"reasoningEffortFormat": "chat-completions",
"modelOptions": {
"temperature": 0.4,
"top_p": 0.95
}
}
],
"settings": {
"qwen/qwen3.8-27b": {
"reasoningEffort": "low"
}
}
}
]How it works
All MCP clients share one browser and their tool calls run in order. Sub-agents and scripts get private browsers on a second engine, so they never disturb yours.
Details, including the tool queue, isolation and crash recovery, are in the architecture notes.
Security
Whoever reaches the MCP endpoint drives a browser on your machine, so the defaults are strict.
127.0.0.1; AUTH_TOKEN adds bearer-token auth for the MCP endpoint and the dashboard.file: and javascript: URLs are refused, and private networks are blocked unless you allow them.no-new-privileges.Pages are untrusted input: text on a page can try to steer a model. Keep a human in the loop for sensitive accounts. Security policy
Clone, build, connect. The setup guide takes you from nothing to a local model browsing in LM Studio in a few minutes.