Odysseus Web MCP — Secure Web Search and Fetch Server for AI Assistants

Odysseus Web MCP is a standalone Model Context Protocol (MCP) server for safe public-web search and URL fetching. It runs locally over stdio and gives MCP-compatible AI assistants two retrieval tools: web_search to discover sources and web_fetch to retrieve and extract public URLs.

Built for clients such as Claude Code, Cursor, and Codex, it combines search-provider fallback, readable HTML/PDF/text extraction, optional JavaScript rendering, and SSRF protections including DNS validation and redirect rechecks.

Features

• Search the public web with provider fallback and ranked, attributed sources.
• Fetch and extract HTML, PDF, and text content from public URLs.
• Protect against SSRF with public-network checks, DNS validation, and redirect revalidation.
• Return bounded, evidence-oriented output with cursors, quality signals, and discovered links.
• Optionally render JavaScript-heavy pages in isolated Playwright.

Install in minutes

Requirements: Python 3.11+ and uv.

# after downloading/extracting this folder (or cloning your copy)
cd odysseus-web-mcp
uv venv .venv
uv pip install -e '.[dev]'
./run-web-mcp.sh

The server communicates over stdio, so it does not open a web port and does not need to be installed into your host application's Python environment. Register the absolute launcher path in your MCP client:

{
  "name": "odysseus-web-mcp",
  "command": "/absolute/path/to/odysseus-web-mcp/run-web-mcp.sh",
  "args": [],
  "cwd": "/absolute/path/to/odysseus-web-mcp"
}

The launcher automatically uses the package's .venv. State defaults to ~/.local/share/odysseus-web-mcp; set WEB_MCP_DATA_DIR to place it elsewhere. No API key is required for the default fallback path, though Brave, Tavily, and Serper keys can be added when you want those providers.

The two tools

web_search

Use it to discover sources for a focused question. It accepts one to three queries plus optional mode, vertical, and freshness controls.

{
  "queries": "Model Context Protocol Python SDK",
  "mode": "discovery",
  "vertical": "general"
}

The response contains ranked URLs, titles, snippets, provider attempts, cache state, a plain-text display projection, and an evidence_id. A host can take any returned URL directly into web_fetch.

web_fetch

Use it to read a known public URL or a bounded batch of URLs.

{
  "url": "https://example.com",
  "focus": "the page's purpose",
  "render": "auto"
}

It returns extracted text, title and document kind, content quality, link discovery, redirect history, HTTP status, truncation/continuation metadata, and an evidence_id. Private and special-use destinations are rejected before transport by default.

Example: how an agent uses the MCP

An agent normally uses the tools as a two-step retrieval loop: search first, then fetch the source it wants to inspect. The payloads below show the shape of a real MCP interaction; IDs and result text are abbreviated for readability.

1. Agent searches for sources

{
  "name": "web_search",
  "arguments": {
    "queries": "official Model Context Protocol architecture",
    "mode": "grounding",
    "vertical": "general"
  }
}

The MCP returns a text content block containing structured JSON:

{
  "status": "ok",
  "query": "official Model Context Protocol architecture",
  "sources": [
    {
      "title": "Architecture - Model Context Protocol",
      "url": "https://modelcontextprotocol.io/docs/concepts/architecture",
      "snippet": "Understand the architecture and communication model...",
      "provider": "duckduckgo",
      "relevance_score": 1.0
    }
  ],
  "provider_attempts": {
    "searxng": "empty",
    "duckduckgo": "ok"
  },
  "evidence_id": "a1b2c3d4...",
  "exit_code": 0
}

2. Agent fetches the selected source

The agent takes the returned URL and calls the second tool:

{
  "name": "web_fetch",
  "arguments": {
    "url": "https://modelcontextprotocol.io/docs/concepts/architecture",
    "focus": "How do clients and servers communicate?",
    "render": "auto"
  }
}

The MCP returns bounded, extracted evidence:

{
  "success": true,
  "url": "https://modelcontextprotocol.io/docs/concepts/architecture",
  "final_url": "https://modelcontextprotocol.io/docs/concepts/architecture",
  "http_status": 200,
  "document_kind": "html",
  "content_quality": "good",
  "content": "The Model Context Protocol defines how clients and servers...",
  "links": [
    {
      "url": "https://modelcontextprotocol.io/docs/concepts/transports",
      "text": "Transports"
    }
  ],
  "evidence_id": "e5f6g7h8...",
  "exit_code": 0
}

The agent can now answer the user from the extracted content, preserve the evidence_id for traceability, and continue with another web_fetch using a returned cursor if the page was longer than the output budget.

How it works locally

MCP host ──stdio──▶ mcp_server.py
                       ├─ web_search → provider chain → ranked evidence
                       └─ web_fetch  → security → HTTP/extract/render → evidence

All persistent state is rooted under WEB_MCP_DATA_DIR. The package has no runtime imports from Odysseus and no access to its credentials, database, memory, browser profiles, scheduler, or agent loop.

Read the full local system design in docs/TECHNICAL_DESIGN.md, and see how the GIFs were recorded in docs/INTERACTIVE_DEMO.md.

Search providers and configuration

The default provider chain is:

SearXNG → Brave → Tavily → Serper → DuckDuckGo → Wikipedia → Bing

Configure it with WEB_MCP_SEARCH_PROVIDER_CHAIN. Optional credentials are DATA_BRAVE_API_KEY, TAVILY_API_KEY, and SERPER_API_KEY. Copy .env.example as a reference, but keep secrets in the host environment rather than committing them.

The optional browser path is disabled by default:

uv pip install -e '.[render]'
./.venv/bin/python -m playwright install chromium
export WEB_MCP_RENDER_ENABLED=true

Distribution and discovery

The server is published in the official MCP Registry under io.github.AceAtDev/odysseus-web-mcp.

For Claude Desktop and other MCPB-compatible clients, download the validated MCPB release bundle from the v0.1.0 GitHub Release. The bundle uses the uv runtime to resolve the declared Python dependencies without shipping a machine-specific virtual environment.

Verify it yourself

The project has a focused test suite and a live qualification runner:

./.venv/bin/python -m pytest -q
./.venv/bin/python tests/live_20_cases.py --output reports/live-20-cases.json

The live qualification runs 10 searches and 10 fetches through the real MCP launcher with disposable state. The latest verification record is in VERIFICATION.md.

To re-record the terminal previews from fresh live calls (requires ImageMagick's convert command):

./.venv/bin/python demos/record_terminal_demos.py

Each GIF is intentionally under ten seconds and shows a real MCP handshake and result shape, not a static product mockup.

Project boundaries

This package is a retrieval primitive, not an agent loop, general-purpose crawler, scheduler, memory store, browser-profile manager, or credential vault. It is designed to be downloaded and connected as an independent MCP server.

License and status

This is the standalone extraction workspace for the Odysseus web search/fetch capability. See MIGRATION_MAP.md for the source-to-module mapping and VERIFICATION.md for the current evidence-based status.