You can ask Codex to work with local PDFs through localipdf's MCP connection. MCP lets an AI assistant discover and call tools. This package uses the same PDF engine as the website and the same server as the Claude Desktop integration. You do not need a second converter.

Before you start

Use a local Codex installation, Node.js 20 or newer, and an installed Chrome, Edge, Chromium or Brave browser. This package needs a computer that can run a local browser; it is not a hosted connector for ChatGPT in a web browser.

No localipdf API key is required. Your normal Codex account and usage arrangements still apply. Start with two disposable PDFs so you can check the connection without sharing sensitive material.

Download the tools

  1. Download the localipdf MCP package.
  2. Extract it into a permanent folder, such as a folder named localipdf-mcp.
  3. Keep server.mjs, driver.js, tools.json, manifest.json, README.md and the site folder together. Moving only the server file will break the connection.

If you already have the extracted Claude package, you can reuse that folder. Both downloads contain the same server. Codex does not need the Claude extension installer.

Add it to Codex

Check your installed client's commands with codex mcp --help. For a CLI installation, replace the example path below with the full path to your extracted server:

codex mcp add localipdf -- node "/absolute/path/localipdf-mcp/server.mjs"

If localipdf is already configured, inspect that entry instead of adding another. If the desktop app cannot find Node, use the full path to the Node executable in place of node.

For longer conversions, add the following setting inside the existing [mcp_servers.localipdf] section of ~/.codex/config.toml:

tool_timeout_sec = 960

This gives the connection time to receive the server's own 15-minute timeout response. It does not make a conversion faster. Keep other server settings intact.

In the desktop app's MCP settings, save the connection and use the available restart control. For CLI, start a fresh session and use /mcp to inspect connected servers. The exact settings labels can vary by client version. See OpenAI's MCP setup documentation for its current configuration and reload steps.

Confirm the connection works

First ask:

List the localipdf tools available through MCP.

Then use two disposable PDFs and an existing output folder:

Use localipdf to merge "/full/path/first.pdf" and
"/full/path/second.pdf" in that order.
Save the result in "/full/path/results".

Open the saved file and check its page count and order. Listing the tools checks the connection. Merging the sample PDFs also checks the browser, file access and saving.

The Merge PDF tool offers the same workflow in the browser if you prefer to choose files yourself.

Choose options before running a tool

Ask Codex to show the available options for the selected tool and file before changing an unfamiliar setting. The connection reads the controls shown by that tool.

You can ask to split a PDF, extract selected pages, convert a document or compress a copy. Give the full file paths and the folder where you want the result saved. The server preserves existing files by choosing a new name when necessary. Check the returned path rather than assuming it replaced an earlier output.

Some operations require your confirmation. Review the action and any warning before agreeing, especially when changing signatures, security restrictions or visible content.

Know which workflows need the website

The tool catalog marks workflows that cannot run through this connection. Signing, drag-based page organization, reading, visual comparison, overlay templates, camera scanning and the workflow builder need the website's interface. The downloadable MCP package also excludes OCR recognition assets.

Use the website for those workflows. They appear in the catalog with an explanation of why they cannot run through the connection.

Understand what stays local

The server processes PDFs in a separate temporary browser profile on your computer. It does not reuse your personal browser tabs or browsing profile.

However, Codex receives your prompts and the tool's responses. These can contain paths, messages and, for some operations, extracted document text. Your AI provider can therefore receive some document information even though the PDF is processed locally. Review your client's data settings before asking it to extract sensitive text.

Read the localipdf privacy policy for the website's practices. Your AI client's policies apply separately.

Update or troubleshoot the connection

To update, extract a fresh download into a new folder, change the configured server path and reload the connection. Test a disposable file before deleting the previous package.

What you see What to check
Node or the server cannot be found Use full executable and server paths; confirm the folder was extracted
A missing file or blank tool page Keep the complete package together, including site and tools.json
No supported browser is found Install a supported browser, or set LOCALIPDF_BROWSER in the server environment to its executable
A conversion times out Check the timeout setting and try fewer pages or a smaller file
The tool is listed but cannot run here Read its limitation and use the website's interface

Questions

Do I need a separate PDF engine for Codex and Claude?

No. Both integrations use the same localipdf server and bundled website tools. Each client has its own connection settings.

Can I install this ZIP as a hosted ChatGPT web connector?

No. This package uses local STDIO and launches a local browser. ChatGPT web needs a remote MCP connection or a supported tunnel, as described in OpenAI's ChatGPT connection guide. That requires a separate connection and access design; it is not part of this setup.

Will the connection change my original PDF?

Results are saved as new files. If an output name already exists, the server chooses another name. Review the saved copy before using or sharing it.

Why should I test with disposable PDFs first?

It confirms that the configured path, browser and output folder work together. It also lets you see what information the assistant receives before using a private document.