Document MCP

ChatGPT-style chatbot plus a Document MCP server. You register folder paths and a local LiteLLM model in Settings. Questions go to LiteLLM with MCP tools; the server reads only those folders, using pywin32 COM for Word / Excel / PowerPoint (same Office path as Model Council).

Chat UI  →  LiteLLM /v1/chat/completions (+ tools)
              ↑ tool_calls (search_documents, read_document, …)
MCP host     →  Document MCP (stdio JSON-RPC)
                    →  allowlisted folders
                    →  win32com Word / Excel / PowerPoint

Claude Desktop, Claude Code, and Cursor attach to the same MCP server and the same folder list.

Quick start

1. Run your LiteLLM proxy.
2. Double-click start.bat.
3. Open http://127.0.0.1:7860.
4. Settings → LiteLLM: paste base URL (e.g. http://localhost:4000), API key, Load models or Add model, then Save.
5. Settings → Folders: paste Windows folder paths. sample_docs is registered on first run.
6. Ask anything those files can answer.

Manual

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
copy .env.example .env
python -m app

Settings

Tab	What it does
LiteLLM	Base URL, key, load /v1/models, add a local model name that is not in the list, test connection
Folders	MCP roots (allowlist). Same idea as Model Council OFFICE_ALLOWED_ROOTS
Claude MCP	Copy Claude Desktop / Claude Code / Cursor snippets, or write Claude Desktop config for you

URLs can be pasted as http://host:4000, http://host:4000/v1, or .../v1/chat/completions — they are normalized like Model Council.

How LiteLLM + MCP work here

Taken from Model Council’s in-app MCP loop:

1. Host calls MCP tools/list and converts schemas to OpenAI tools.
2. Host POSTs chat/completions to LiteLLM with those tools (tool_choice: auto).
3. If the model returns tool_calls, the host runs MCP tools/call (stdio).
4. Tool JSON is appended as role: tool and the loop repeats until a normal answer.

The model never sees files unless it calls a tool. Roots are allowlisted. Office files prefer COM so Windows-unlocked / IRM files can still be read.

MCP tools

Tool	Purpose
ping	Health, engines, folder count
list_folders	Registered roots
add_folder / remove_folder	Change roots
list_documents	Inventory
search_documents	Name + extracted text
read_document	Extract (COM first, library fallback)

Prompt: ask_documents.

Claude Desktop / Claude Code / Cursor

In the chatbot: Settings → Claude MCP → Add to Claude Desktop (writes %APPDATA%\Claude\claude_desktop_config.json). Restart Claude, then ask it to ping Document MCP.

Or copy this shape (paths are filled in by Settings):

{
  "mcpServers": {
    "document-mcp": {
      "command": "D:\\\\AntigravityCodes\\\\Document-MCP\\\\.venv\\\\Scripts\\\\python.exe",
      "args": ["-m", "document_mcp"],
      "cwd": "D:\\\\AntigravityCodes\\\\Document-MCP",
      "env": {
        "PYTHONUTF8": "1",
        "DOCUMENT_MCP_ROOTS": "D:\\\\AntigravityCodes\\\\Document-MCP\\\\data\\\\folders.json"
      }
    }
  }
}

Claude Code:

claude mcp add-json document-mcp "{...stdio server json from Settings...}"

Cursor: repo mcp.json, or Settings → Claude MCP copy the Cursor snippet.

HTTP (LiteLLM MCP gateway):

start_mcp_http.bat

Then add http://127.0.0.1:8765/mcp.

pywin32 / Office

• Word.Application → .doc / .docx
• Excel.Application → .xls / .xlsx / .xlsm
• PowerPoint.Application → .ppt / .pptx

Fallback: python-docx / openpyxl / python-pptx. PDFs: PyMuPDF. Legacy binary Office still needs Microsoft Office.

python -m document_mcp.reader sample_docs\mena_rhq_overview.md

Layout

app/                 FastAPI chat host (MCP client + UI)
document_mcp/        FastMCP server, COM readers, Claude install helper
sample_docs/         Starter files
data/                folders.json, settings, extract cache (local)