Desktop Commander with Codex and Claude for Small Busin...

Most small businesses already pay for an AI assistant. Fewer have connected it to the place the work actually lives: the folder of supplier invoices, the spreadsheet exported from the till, the website files, the server that runs the booking page. Desktop Commander is a free, open-source bridge that lets Claude or Codex read and change those files and run commands on a computer you choose. This guide covers what it does, how to install it for both assistants, what to lock down first, and ten ordinary jobs worth trying.

It is the hands-on companion to Desktop Commander, Cowork, and Browser Agents for a Small Business, which compares it with task workspaces and browser assistants. Product details below were checked on October 14, 2026, when version 0.3.0 was current.

What it actually gives an assistant

Desktop Commander is an MCP server. MCP is the plug standard that lets an assistant use outside tools. Once connected, the assistant can ask Desktop Commander to do things, and Desktop Commander does them on that computer with your user account's permissions.

Scroll to compare columns and rows when the table extends beyond the box.

Tool group	In plain words
Terminal and processes	Run commands, keep a long-running program open (a database prompt, an SSH session, a local preview server), read its output, stop it
Files and folders	Read, write, move and list files; read the end of a long log; get file details
Search	Find files by name or by what is written inside them, across a folder tree
Targeted edits	Replace one passage in a file instead of rewriting the whole thing
Office files	Read, write and search Excel and Word files; extract text from PDFs, build a PDF from Markdown, add or remove PDF pages
Code without files	Run a short Python, Node.js or R calculation in memory
Remote Device	Let ChatGPT or Claude on the web reach the computer while a local launcher is running

It is MIT-licensed and free. Its README says it works with an ordinary Claude subscription without API charges. The assistant subscription is the cost, not the bridge. Desktop Commander README.

Do Codex and Claude Code even need it?

Claude Code and the Codex CLI already run in a terminal and already have their own command and file tools. If that is where you work, Desktop Commander is optional. It still adds three things those tools do not do the same way:

Office documents. Reading and editing Excel and Word files, and pulling text out of PDFs, are built in.
Sessions that stay open. An interactive database prompt or SSH session can stay running while the assistant reads and responds to it.
One toolset everywhere. The same tools, settings and limits work in Claude Desktop, Claude Code, the Codex CLI, the ChatGPT desktop app and the Codex IDE extension.

The assistants that gain most are the chat apps without their own terminal: Claude Desktop, and ChatGPT on the web through the Remote Device. For an owner who never opens a terminal, that is the whole point.

Set it up for Claude

Claude Desktop. Add an entry to claude_desktop_config.json (on Windows, %APPDATA%\Claude\claude_desktop_config.json; on a Mac, ~/Library/Application Support/Claude/claude_desktop_config.json), then restart Claude:

{
  "mcpServers": {
    "desktop-commander": {
      "command": "npx",
      "args": ["-y", "@wonderwhy-er/desktop-commander@0.3.0"]
    }
  }
}

The README's examples use @latest, which updates itself on every restart. Pin the version you tested and update deliberately, the same way you would any tool that can change files. It needs Node.js installed. The README also offers a one-line setup command and a Docker image.

Claude Code. One command, available in every project with --scope user (leave the flag off to install it for the current project only):

claude mcp add --scope user desktop-commander -- npx -y @wonderwhy-er/desktop-commander@0.3.0

Claude Code asks before using new tools. Permission rules name Desktop Commander's tools as mcp__desktop-commander__<tool>, so you can deny one specifically. Claude Code MCP documentation and permission rules.

Set it up for Codex

codex mcp add desktop-commander -- npx -y @wonderwhy-er/desktop-commander@0.3.0

Or edit ~/.codex/config.toml by hand, which lets you set timeouts and switch individual tools off:

[mcp_servers.desktop-commander]
command = "npx"
args = ["-y", "@wonderwhy-er/desktop-commander@0.3.0"]
startup_timeout_sec = 30
tool_timeout_sec = 120
disabled_tools = ["set_config_value"]

OpenAI's documentation says the ChatGPT desktop app, the Codex CLI and the IDE extension share this configuration, so one entry serves all three. codex mcp list shows what is configured. Codex MCP documentation.

Lock it down before the first real job

The README is candid: these settings are guardrails, not a sandbox. Plan around that.

Give it a work folder, not your whole disk. Set allowedDirectories to a folder such as AI-Work. An empty list means the entire filesystem. The limit applies only to Desktop Commander's file tools. A terminal command can still reach other folders, so the folder is a habit, not a wall.
Block commands you never want run. blockedCommands takes a list. Treat it as a speed bump: the README notes that command substitution, symlinks and running code can get around it.
Change settings in a separate chat. The README warns that an assistant blocked by allowedDirectories may try to change that setting itself. Configure first, then switch off the set_config_value tool, as in the Codex example above, or with a Claude Code deny rule for mcp__desktop-commander__set_config_value.
Turn off telemetry if you prefer. It is limited and pseudonymous, and telemetryEnabled: false stops it.
Know what the logs keep. Desktop Commander logs tool names and their arguments, unredacted, in its settings folder. Do not paste passwords or card numbers into a request; they would end up in that log.
Work on copies until you trust a workflow. Copy the invoices or the export into the work folder. Keep the originals where they are, and back them up.
Use the Docker install for anything sensitive. The README recommends it for stronger isolation from the rest of the computer.

The Remote Device: useful, and worth respecting

npx @wonderwhy-er/desktop-commander@latest remote starts a launcher on your computer. After a browser sign-in, you connect ChatGPT or Claude on the web through mcp.desktopcommander.app. Commands run on your computer under your account, and only while the launcher is running; Ctrl+C stops it.

That makes it a good fit for "run this on the office server while I am out" and a poor fit for a launcher left running unattended on a machine with customer records. Start it for a job, watch it, stop it. Remember that the relay is part of the path: the command runs locally, but the conversation does not become local. The earlier comparison covers that data path in more detail.

Ten jobs worth handing it

Each is written as something you could type. Start each one in the work folder.

Invoices to a spreadsheet. "Read every PDF in AI-Work/invoices-september, and make september-invoices.xlsx with vendor, invoice number, date, total and tax. List any PDF you could not read." Check: open three invoices and compare.
Clean a till or online-store export. "In sales-export.csv, remove exact duplicate rows, total sales by category and by week, and save the summary as a new Excel file. Do not change the original." Check: the grand total matches the till report.
Compare two price lists. "Compare supplier-prices-old.xlsx and supplier-prices-new.xlsx by item code and list every price that rose more than 5 percent." Check: spot-test five rows.
Find the document you need. "Search AI-Work/contracts for any file that mentions automatic renewal, and tell me the file names and the renewal dates." Check: open the matches.
Tidy a folder, with a dry run first. "Propose new names for the photos in listing-photos using the date and product name, as a table. Do not rename anything until I approve." Then approve or adjust. Check: a few files before and after.
Draft a document from a template. "Using quote-template.docx, create a quote for the customer details in quote-request.txt." Check: every price and name, before it goes anywhere.
Update the website's hours or prices everywhere. "In the website folder, find every page that shows our Saturday hours, change them to 9 to 3, and show me each change before saving." Check: build or preview the site before publishing.
Read a log you do not understand. "Read the last 200 lines of booking-app.log and explain the errors in plain English, with what you would check first." Check: the explanation matches the time of the problem.
Back up a folder and prove it. "Copy AI-Work/finance-2026 to the backup drive, then compare checksums of every file and report any mismatch." Check: the report names a file count you expect.
A health check on a small server. "Report free disk space, memory, any failed services and the date of the last backup log entry. Change nothing." Check: read-only means read-only; stop if it proposes a change.

Notice the pattern: a named folder, a new output instead of an overwritten original, and one check a person does. That is what keeps a helpful tool from becoming a confident mistake.

Claude or Codex for which job?

Both work with the same tools, so start with the subscription you already pay for.

Claude Desktop is the friendliest for an owner who wants to type requests and review results in a chat window.
Claude Code or the Codex CLI suit repeatable work you will run again next month, because the steps can be saved as a script.
The ChatGPT desktop app reaches the same Desktop Commander configuration as Codex, so a team already in ChatGPT does not need a second setup.

Do not point two assistants at the same files at the same time. Finish one job, check it, then start the next.

What not to hand it

Keep payments, banking, payroll, tax filings and legal decisions with a person and the systems built for them. Do not let it delete anything that is not backed up. Do not move customer personal data into new places, including new spreadsheets, without a reason you could explain to the customer. Use it to prepare and check work, and approve each change that matters.

A one-week start

Day 1: Install it for the assistant you already use. Make the work folder. Set allowedDirectories, telemetry and blocked commands, then disable set_config_value.
Day 2: Run job 4, a search. It changes nothing and shows how the assistant reports.
Days 3 to 4: Run job 1 or 2 on copies and check the totals by hand.
Day 5: Write down what saved time and what needed correcting. Keep the jobs that passed; drop the rest.

Fact-check notes and sources

Desktop Commander capabilities, installation, configuration keys, log behavior, telemetry, Remote Device and security warnings: Desktop Commander README, checked October 14, 2026. Version 0.3.0 was published to npm on October 13, 2026.
Codex configuration and the shared ChatGPT app, CLI and IDE extension config: Codex MCP documentation.
Claude Code server scopes and permission rule names: Claude Code MCP and Claude Code permissions.
The ten jobs are examples to adapt, not measured case studies. Results depend on file quality and the model you use.

Desktop Commander with Codex and Claude: A Small Business Owner's Field Guide