site-context-mcp

Problem

Agents invent project and site claims without citations. When you ask what a public site says about a product page, a topic note, or an essay, you often get fluent prose and no way to verify it against the live corpus.

Related MCP server: DuckDuckGo MCP Server

Solution

A stdio MCP server that:

1. Fetches public pages starting from llms.txt, then follows same-host links to core pages

2. Caches them in-process (~1-hour TTL)

3. Exposes tools that return structured answers with a top-level sources[] of public URLs on every response

No secrets, no LinkedIn scraping, no private admin / WOD-APP links — only HTTPS to the wired public site host.

This package is for site/project context. It does not duplicate resume identity or resume-summary tools (see sibling resume-mcp for that).

The default corpus is YongBo Yu’s public site. The same fetch/cache/source-URL shape can be pointed at any public llms.txt-indexed site corpus.

Requirements

• Python 3.11+

• Network access to the wired public host (default: yongbo-yu.vercel.app) when running the live server

Install

# from this repo
pip install -e .

# or with uv
uv pip install -e .

Dev / smoke tests (mocked HTTP; no live network required):

pip install -e ".[dev]"
pytest -q

Run (stdio)

# module entrypoint
python -m site_context_mcp

# console script (after install)
site-context-mcp

# with uv
uv run python -m site_context_mcp

The process speaks MCP over stdin/stdout. Do not pipe unrelated stdout into the same process.

Cursor mcp.json

Add a server entry (Cursor: Settings → MCP, or edit ~/.cursor/mcp.json / project .cursor/mcp.json):

{
  "mcpServers": {
    "site-context-mcp": {
      "command": "python3",
      "args": ["-m", "site_context_mcp"],
      "cwd": "/absolute/path/to/site-context-mcp"
    }
  }
}

If the package is installed into a venv, point command at that interpreter (or use the site-context-mcp console script).

With uv:

{
  "mcpServers": {
    "site-context-mcp": {
      "command": "uv",
      "args": ["run", "python", "-m", "site_context_mcp"],
      "cwd": "/absolute/path/to/site-context-mcp"
    }
  }
}

After saving, reload MCP servers in Cursor. The tools below should appear as site-context-mcp tools.

Tools

Every tool response is JSON text that includes a top-level sources array of public URLs. Partial fetch failures are reported in cache_notes when present.

Example prompts (in Cursor / Claude)

• “List the public pages this MCP has cached and cite the source URLs.”

• “Get /projects/kilodock and quote the engineering metrics with sources.”

• “Search the site corpus for ‘LangGraph’ / ‘Codex’ and return snippets with URLs.”

• “Use get_project for KiloDock evidence — do not invent stack claims.”

Data sources (public only)

On startup (and on first tool use if needed), the server fetches:

1. https://yongbo-yu.vercel.app/llms.txt

2. Same-host pages linked from llms.txt (home, about, project pages, topic pages, essays)

3. Optionally resume.json when linked — used only for light project cross-links, not as a resume clone

Binaries and off-host / private targets are skipped. No API keys.

Author

YongBo Yu (also Yong Yu) — Toronto, Canada · GitHub YongBoYu1

• Site: https://yongbo-yu.vercel.app

• KiloDock evidence: https://yongbo-yu.vercel.app/projects/kilodock

• Site index: https://yongbo-yu.vercel.app/llms.txt

License

MIT — see LICENSE.