Document Organizer MCP

This is a starter MCP server you can plug into an agent that needs to do more with documents than basic chat alone.

It gives your agent a few practical abilities:

• list documents inside a chosen folder

• inspect document metadata and previews

• extract text from common file types

• search across many documents at once

• summarize the overall collection

• propose cleaner filenames from document content

• move files safely inside the allowed document root

• optionally search and read Google Drive files with an OAuth token

Why this is a good first custom MCP

You said you have never built one before, so this project stays small on purpose:

• one server file

• one configurable document root

• a handful of tools that are easy to test

• no database, auth flow, or cloud deployment required for version one

Once this works, you can keep adding richer behaviors like tagging, rename suggestions, duplicate detection, filing rules, Drive integration, or OCR.

Important for ChatGPT Business

ChatGPT Business can connect only to a remote MCP server, not one running only on your laptop or inside this workspace.

That is why this project now includes remote hosting files:

• Dockerfile

• render.yaml

• requirements.txt

This starter is set up for a simple first deployment without authentication. That is the easiest way to get unstuck, but it also means:

• only upload non-sensitive test documents at first

• keep documents inside the data/ folder

• do not point DOC_ROOT at your whole machine or workspace

Once you are comfortable, the next upgrade should be adding authentication.

Supported file types

• .txt

• .md

• .json

• .csv

• .pdf

• .docx

• .xlsx

Project layout

document-organizer-mcp/
├── pyproject.toml
├── README.md
└── server.py

Setup

From this folder:

cd /workspace/document-organizer-mcp
uv sync

Pick the folder your agent should be allowed to inspect. For example:

export DOC_ROOT=/workspace/document-organizer-mcp/data

Optional: restrict which extensions are searchable.

export DOC_EXTENSIONS=.pdf,.docx,.md,.txt

Run locally

For local development with the MCP inspector:

uv run mcp dev server.py

For direct stdio use:

uv run python server.py

For the remote HTTP shape that ChatGPT expects:

HOST=0.0.0.0 PORT=8000 uv run python server.py

The MCP endpoint will be:

http://localhost:8000/mcp

Available tools

healthcheck

Returns the active document root and allowed extensions.

list_documents

Shows files under the document root.

Example:

{
  "subdir": "contracts",
  "limit": 25
}

inspect_document

Returns metadata and a short preview for a single file.

extract_document_text

Pulls document text into the agent context.

search_documents

Searches across documents and returns snippets around each hit.

summarize_collection

Returns collection-level stats so the agent can get oriented quickly.

propose_rename

Suggests a better filename using the first meaningful line of extracted text and a detected date when available.

move_document

Moves a file inside DOC_ROOT. It defaults to dry_run=true, which is the safer way to start.

Example:

{
  "source_path": "inbox/scan-001.pdf",
  "destination_dir": "contracts/2026",
  "new_name": "2026-05-22 Master Services Agreement.pdf",
  "dry_run": true
}

drive_status

Checks whether the optional Google Drive bridge has been configured.

drive_search_files

Searches Google Drive through the Drive API. The query field uses normal Drive API query syntax.

Example:

{
  "query": "name contains 'invoice' and 'root' in parents",
  "page_size": 10
}

drive_get_file

Fetches metadata for a single Drive file.

drive_extract_text

Exports or downloads text from Drive files. For Google Docs and Slides it uses plain text export. For Google Sheets it uses CSV export of the first sheet.

search

A ChatGPT-friendly search entry point. It returns search results in the shape OpenAI documents for MCP search tools.

fetch

A ChatGPT-friendly fetch entry point. It returns the full contents of a selected search result item.

Google Drive setup

The Drive bridge is optional. It uses a bearer token in:

export GOOGLE_DRIVE_ACCESS_TOKEN=your_access_token_here

For a first build, the simplest path is:

1. Create a Google Cloud project.

2. Enable the Google Drive API.

3. Create OAuth client credentials.

4. Generate a user access token with a Drive scope such as drive.readonly or drive.file.

5. Pass that access token into the MCP host as GOOGLE_DRIVE_ACCESS_TOKEN.

This starter keeps auth intentionally simple so you can get the bridge working before you build token refresh or a full sign-in flow.

Deploy to Render

This is the easiest path I recommend for a first deployment.

1. Put only safe test files in data/

Use this folder:

document-organizer-mcp/data/

Do not deploy your full workspace as a public no-auth MCP.

2. Put this project in a GitHub repo

Render deploys most smoothly from GitHub. Create a repo and upload the contents of document-organizer-mcp/.

3. Create a new Render Web Service

In Render:

1. Click New +

2. Choose Web Service

3. Connect your GitHub repo

4. Render should detect the included Dockerfile

4. Set environment variables in Render

Use these values:

• HOST = 0.0.0.0

• PORT = 8000

• DOC_ROOT = /app/data

Optional:

• DOC_EXTENSIONS = .pdf,.docx,.md,.txt,.xlsx,.csv,.json

• GOOGLE_DRIVE_ACCESS_TOKEN = your temporary OAuth access token if you want Drive read access

5. Deploy

After deploy finishes, your MCP endpoint will look like:

https://your-service-name.onrender.com/mcp

Save that URL. That is what ChatGPT Business will connect to.

Add it in ChatGPT Business

As of May 27, 2026, ChatGPT Business uses developer mode and remote MCP apps for this flow.

1. Enable developer mode

In ChatGPT web:

1. Open Workspace Settings

2. Go to Permissions & Roles

3. Find Connected Data Developer mode / Create custom MCP connectors

4. Turn it on for yourself

On Business, only admins or owners can do this.

2. Create the app

In ChatGPT web:

1. Open Workspace Settings

2. Go to Apps

3. Click Create

4. Choose the option to add a custom MCP app

5. Paste your remote MCP URL, for example:

https://your-service-name.onrender.com/mcp

3. Test privately

Start a new chat and attach or invoke the draft app. Try prompts like:

• Summarize the document collection available in this app

• Find documents about invoices

• Suggest cleaner names for the available files

4. Publish

When it looks good:

1. Go back to Workspace Settings

2. Open Apps

3. Find the draft

4. Click Publish

Business note:

• published apps currently are not edited in place

• if you change tools or metadata later, recreate and republish

How I would use this with your document organizer agent

A good first workflow for the agent is:

1. summarize_collection to understand the folder.

2. list_documents to discover likely targets.

3. inspect_document or extract_document_text to read them.

4. propose_rename to produce cleaner names.

5. move_document with dry_run=true first.

6. move_document with dry_run=false once the plan looks right.

7. drive_search_files and drive_extract_text when you want the same agent to reach into Drive.

8. search and fetch when ChatGPT wants a connector-style search/fetch flow.

Good next upgrades

If you want your document organizer agent to become genuinely useful, these are the next features I would add:

1. suggest_folders_for_document Use file content plus filename patterns to recommend where a document belongs.

2. find_duplicates Detect repeated files by hash, name similarity, or matching extracted text.

3. tag_document Save lightweight metadata in a sidecar JSON file.

4. index_cache Cache extracted text so large collections search faster.

5. oauth_refresh_flow Replace the one-off Drive access token with a renewable sign-in flow.

6. drive_move_or_label Extend the Drive bridge from read-only access into organization actions.

Notes on the MCP stack

This starter uses the official Python MCP SDK and FastMCP, which is the current stable path in the official Python SDK documentation as of May 27, 2026.