- Blog
- Bookkeeping Automation
- Claude Code vs Codex for Bookkeeping: 4 AI Work Surfaces Compared
Claude Code vs Codex for Bookkeeping: 4 AI Work Surfaces Compared
title: "Claude Code vs Codex for Bookkeeping: 4 AI Work Surfaces Compared" meta: title: "Claude Code vs Codex for Bookkeeping: 4 Tools Compared" description: "Claude Code, Claude (formerly Cowork), the ChatGPT app with Codex, and Codex CLI, scored for bookkeeping work as of September 2026. Plus setup steps."
Claude Code vs Codex is the question most bookkeepers ask first, but there are really four tools on the table. Claude Code. Claude, which just absorbed Cowork. The ChatGPT desktop app, which now carries Codex inside it. And Codex CLI. They share a lot of vocabulary now. Skills. Connectors. Agents. So which one should touch a client's books?
Our take, up front: for serious bookkeeping work, Claude Code and the command-line tools come out ahead. They work on your real files, they run your written procedures the same way every time, and you can watch every step. But the tool is not the moat. Two things matter more than which logo you pick. You need good accounting skills, meaning written instructions the tool follows. And you need your books connected through MCP or a connector. Get those two right and any of the four gets better. Skip them and none of them will save you much time.
One scope note. This page compares four places you do the work. It doesn't compare the models underneath. If you want Claude vs ChatGPT as assistants, read Claude vs ChatGPT for bookkeeping and ChatGPT vs Claude for accounting. This one is about the work surface: a terminal agent, a cloud agent, a desktop app, and a command-line tool.
Which is better for bookkeeping: Claude Code, Codex, or the Claude and ChatGPT apps?
As of September 2026, Claude Code scores highest in our weighting (4.55 out of 5), followed by Codex CLI (4.30), the ChatGPT desktop app with Codex mode (4.15), and Claude with its former Cowork tasks (3.80). The two terminal tools win on working with real files, running repeatable skills, and letting you review each step. The apps win on ease of starting. Either way, the tool matters less than two things you set up yourself: a written accounting skill the tool follows, and a live connection to your books through MCP or a connector. Start that connection read-only.
What changed in September 2026
On September 16, 2026, Anthropic folded Cowork into Claude. Its pricing page now says: "Claude Cowork is now just Claude. Rolling out to Pro and Max, with more plans to follow." Press coverage reports that chats, projects, artifacts, connectors, and skills carry over. So if you searched for Claude Cowork, it's still here. It's just called Claude now, and Cowork-style tasks run inside regular conversations. On the OpenAI side, what people called the ChatGPT Codex app now shows up in OpenAI's help center as the ChatGPT desktop app's Codex mode. Codex CLI is still its own tool.
Key Takeaways
- Claude Code scores highest for bookkeeping work: 4.55 of 5 in our weighting, driven by local files, skills, and one-command MCP setup.
- Codex CLI is a close second: 4.30, with the same terminal strengths and a cheaper start, since Codex comes with ChatGPT plans, including Free and Go.
- The ChatGPT desktop app is the easiest start: 4.15, with chat, Codex mode, and QuickBooks connection paths in one place.
- Claude's former Cowork runs jobs while you're away: 3.80, because cloud tasks keep going with the laptop closed, but you review results after, not step by step.
- Skills are the real advantage: a written bookkeeping procedure makes any of the four more consistent. Without one, you re-explain the job every session.
- Connect your books, read-only first: MCP servers and connectors let the tool see real data. Turn off write access until you trust the workflow.
The four work surfaces, in plain terms
Categorizes the routine. Flags what needs you.
See Growthy on a sample book. Read-only bank access.
Get startedHere's what each one is, as of September 2026.
Claude Code is Anthropic's agent that runs in your terminal. It also runs in an IDE, the desktop app, and the web. It reads the folder you point it at, edits files, and runs commands. You write a skill once (a folder of instructions built around a SKILL.md file) and it follows it every session. You plug in outside tools through MCP servers. A CLAUDE.md file holds the project's memory. For bookkeeping, "the project" is usually one client's folder.
Claude, formerly Cowork, is the Claude app with task mode built in. Per Anthropic's help center, Cowork tasks run in the cloud (in beta) on Anthropic's servers, inside an isolated environment. Claude plans the work, breaks it into subtasks, runs code and shell commands in that environment, and hands back the output. Your sessions and files save to your Claude account. The work keeps going when you close your laptop. When a task needs a local file or your browser, Claude reaches it through the Claude Desktop app, which has to stay open.
The ChatGPT desktop app now does two jobs. There's the chat you already know. And there's Codex mode, which used to be the separate Codex app. That app runs on macOS and Windows. It manages several agents at once, runs Automations (scheduled work), uses worktrees, and supports reusable skills. An April 2026 update added computer use, in-app browsing, image generation, memory, and plugins. OpenAI also lists ChatGPT Work as a separate experience for longer research and deliverable tasks. We don't score Work here.
Codex CLI is OpenAI's terminal tool. It inspects, edits, and runs code from the command line. OpenAI's own description: "Package repeatable instructions as skills, then add plugins to connect Codex to your team's tools and data." You can sign in with your ChatGPT account, which uses your plan's usage, or with an API key, which bills per token.
How we weighted them: the scorecard
Bookkeeping isn't coding, so we didn't score these like coding tools. We picked five things that decide whether a tool helps on client work, weighted by how much each costs you when it's missing.
The five criteria and their weights:
- Works on your actual files (25%): bank CSVs, exports, statements, workpapers, across many clients.
- Connects to your books (20%): how directly it reaches QuickBooks Online through MCP or a connector.
- Runs your skills (20%): whether it follows a written procedure the same way every time.
- Lets you review before anything sticks (20%): whether you see each step and each change before it counts.
- Easy and cheap to start (15%): learning curve plus the plan you need.
Each tool gets 1 to 5 on each criterion. Multiply by the weight, then add. Here's how they land.
Claude Code: 4.55
- Files: 5 (reads and edits the folder you point it at)
- Books: 5 (MCP support built in; our Claude guide adds QuickBooks in one command)
- Skills: 5 (SKILL.md skills plus CLAUDE.md memory per client folder)
- Review: 5 (every command and file change happens in front of you)
- Start: 2 (a terminal is a real learning curve)
- Math: 1.25 + 1.00 + 1.00 + 1.00 + 0.30 = 4.55
Codex CLI: 4.30
- Files: 5 (inspects, edits, and runs things in your folders)
- Books: 3 (plugins connect tools and data, but QuickBooks isn't among the plugin examples OpenAI lists)
- Skills: 5 (skills are a named, first-class feature)
- Review: 5 (runs in your terminal where you see each step)
- Start: 3 (still a terminal, but Codex comes with ChatGPT plans, including Free)
- Math: 1.25 + 0.60 + 1.00 + 1.00 + 0.45 = 4.30
ChatGPT desktop app with Codex mode: 4.15
- Files: 4 (Codex mode works on files; chat is lighter)
- Books: 4 (several QuickBooks connection paths exist on the ChatGPT side)
- Skills: 4 (reusable skills in Codex mode)
- Review: 4 (desktop app with worktrees on its feature list)
- Start: 5 (familiar app, and Codex is included on Free and Go)
- Math: 1.00 + 0.80 + 0.80 + 0.80 + 0.75 = 4.15
Claude, formerly Cowork: 3.80
- Files: 4 (runs code in a cloud environment; local files go through Claude Desktop)
- Books: 4 (connectors reportedly carry over into the new Claude)
- Skills: 4 (skills reportedly carry over too)
- Review: 3 (the job runs while you're away, so you check the result, not each step)
- Start: 4 (it's the Claude app; Pro starts at $17 a month billed annually)
- Math: 1.00 + 0.80 + 0.80 + 0.60 + 0.60 = 3.80
These scores are our judgment against the facts above, not a benchmark. The gaps are small. Say you never want to open a terminal. Set the weights to 20% files, 15% books, 20% skills, 15% review, and 30% start. Now the ChatGPT app moves to first at 4.30, and Claude Code and Codex CLI tie at 4.10. That's the point of showing the math. Re-weight it for your own work.
Claude Code for bookkeeping: the strongest pick
In the Claude Code vs Codex matchup, Claude Code wins by a nose, and it wins because bookkeeping lives in files. Bank statements. CSV exports. A chart of accounts. Last month's workpapers. Claude Code works right in that folder. It can read twelve months of statements, write a cleaned CSV, and show you exactly what it changed.
The CLAUDE.md file matters more than it sounds. Put one in each client's folder. Write down the client's quirks: which account the owner's personal card hits, which vendor is really a contractor, which deposits are loan proceeds. Claude Code reads that file each session. You stop re-explaining the client every month.
Skills are the second half. A skill is a folder of instructions built around a SKILL.md file. Write one for your month-end steps, one for bank statement cleanup, one for your review checklist. Our guide to the Claude bookkeeping skill covers what a skill is and where it stops.
Then connect the books. Claude Code supports MCP servers. Our guide to connecting QuickBooks Online to Claude walks through the setup for both Claude Code and Claude Desktop, including how to keep it read-only.
What it costs: As of September 2026, per Anthropic's pricing page, Claude Code is included on paid Claude plans. Pro is $17 a month billed annually ($200 up front) or $20 billed monthly. Max starts at $100 a month with 5x or 20x Pro usage. Claude Code shares usage with your Claude plan, so heavy terminal days draw from the same pool as chat.
Where it falls short: the terminal. If you've never used one, the first week is slow. That's the whole reason it scores 2 on starting.
Want to go deeper? Read how to use Claude for bookkeeping and Claude for accounting, then try it on files with Claude for bank statement bookkeeping. If you like plain-text books, Claude plus Beancount shows what a terminal agent can do with a ledger that lives in files.
Claude, formerly Cowork: best for long jobs while you're away
The new Claude is built for handing off a job and walking away. Tasks run in the cloud. Claude breaks the work into subtasks, runs code in an isolated environment, and saves the output to your account. Close the laptop and it keeps going.
That fits certain bookkeeping jobs well. A long reconciliation prep across many statements. A first pass at a close checklist. A research task you'll read tomorrow. Our pieces on Claude Cowork for accounting and bookkeeping and Cowork for month-end close cover those workflows. They still apply, since Cowork tasks now run inside Claude.
The trade-off is review. When a job runs while you're gone, you check the finished output instead of each step. For client books, that's a real difference. It's why this surface scores 3 on review. Fix that by writing your review rules into a skill, so the task leaves a trail you can check.
Local access has a catch. A task that needs a file on your computer or your browser goes through the Claude Desktop app, which has to stay open. So plan local-file jobs for when your machine is on. For the QuickBooks side, start with our Claude connection guide, and confirm the connection your task needs is one the cloud side can reach.
For firms: Claude's Team and Enterprise plans list organization-wide skills deployment and admin controls for remote and local connectors. One approved skill, pushed to everyone. One place to decide which connectors staff can use. More on that in Claude Cowork for accounting firms.
The ChatGPT desktop app and Codex mode: the easiest place to start
If you already live in ChatGPT, this is the lowest-friction option. Chat handles questions. Codex mode handles work on files. Both sit in one app on macOS or Windows.
Codex mode has some features bookkeepers can use right away. Automations run scheduled work, so recurring work can run without you starting it. Reusable skills let you save a procedure once. Worktrees are on the feature list too.
The connection story: on the ChatGPT side, there are several ways to connect QuickBooks Online. They include an official path, a third-party connector, and a build-it-yourself route. Each one reaches a different slice of the books. Our guide to connecting QuickBooks Online to ChatGPT compares all three paths. Read it before you pick, because some of those paths can now write, not just read.
What it costs: as of September 2026, OpenAI says Codex is included with ChatGPT plans, including Free and Go. Usage limits vary by plan. Per OpenAI's Codex pricing page, Pro starts at $100 a month with 5x or 20x more Codex usage than Plus.
Where it falls short: the app blends chat and agent work. That makes it easy to ask a chat question against live books with no written procedure behind it. For client work, use Codex mode with a skill, not a one-off chat prompt.
For the chat side on its own, see ChatGPT for bookkeeping and our list of ChatGPT prompts for bookkeepers.
Codex CLI: the terminal option on the OpenAI side
Codex CLI is the closest match to Claude Code. It works in your terminal. It inspects, edits, and runs things in your folders. It treats skills as a core feature. If your firm already pays for ChatGPT, it's the cheapest way to get a terminal agent, since Codex comes with those plans.
You have two ways to sign in. Sign in with ChatGPT and it uses your plan's usage. Or use an API key, which bills at API rates. The API-key route skips the cloud features, like GitHub code review and Slack. For a bookkeeper working locally, that may not matter.
The connection story is the weak spot. Codex connects to outside tools through plugins. The examples OpenAI lists in its plugin marketplace are Gmail, Google Calendar, Google Drive, GitHub, and Documents. QuickBooks isn't among them. So before you plan a workflow around live books in Codex CLI, check the marketplace yourself. Our ChatGPT connection guide covers the OpenAI-side paths into QuickBooks, and our QuickBooks integrations hub collects every connection option we've written up.
That's the only reason Codex trails in the Claude Code vs Codex scoring. On files, skills, and review, the two tie.
The tool isn't the moat
Swap Claude Code for Codex CLI and a good workflow keeps working. Swap in a bad workflow and neither tool helps. Two things carry across all four surfaces.
1. Good accounting skills. A skill is your procedure, written down. It covers what to check first, how to handle a transfer, when to stop and ask. Anthropic and OpenAI both now call these skills. Write yours once and the tool follows it the same way every time. Without one, you're typing the same instructions into every session and getting slightly different results. Start with our Claude bookkeeping skill guide. A caution: a skill written for Claude may need changes before it runs in Codex. Test it on the other side before you trust it.
2. Your books, connected. A tool that can't see the books is working from whatever you paste in. MCP servers and connectors fix that. Claude Code and Claude use MCP servers and connectors. The ChatGPT app has its own connection paths. Codex CLI uses plugins. The setup differs, but the rule doesn't: connect the real data, and start read-only. Our guides for Claude and ChatGPT cover the steps, and the QuickBooks integrations hub has the rest.
Skills plus connected books is the setup. The tool is just where you run it.
Which one should you pick?
Start with who you are, not which tool is newest.
If you're a solo bookkeeper (a handful of clients up to 30 or so):
- Pick Claude Code if you're willing to spend a week learning the terminal. One folder per client, a CLAUDE.md in each, and one month-end skill. That setup pays you back every close.
- Pick Codex CLI instead if you already pay for ChatGPT and don't want a second subscription. Check the QuickBooks connection path first.
- Fallback: if the terminal is a hard no, use the ChatGPT desktop app in Codex mode, or Claude's task mode. You give up some review control. You gain an easier start.
- First week: write the skill before you connect any client's books. Test it on one client's exported CSVs.
If you run a small firm (a few staff, many clients):
- Pick one terminal tool for the person who builds your procedures. That person writes and tests the skills.
- Give everyone else the app. On Claude, Team and Enterprise plans list organization-wide skills deployment and admin controls for connectors. That's how one tested skill reaches the whole team.
- Decide connector access centrally. Which staff can connect which client books, and whether any of them get write access.
- First month: standardize one workflow, like bank statement cleanup, before you add a second. Bobby's newsletter, Scaling Clients, covers this kind of firm-systems work for bookkeeping and accounting firm owners.
If you're an owner doing your own books:
- Pick the app you already use. For most owners, that's the ChatGPT desktop app or Claude. Codex comes with ChatGPT plans, including Free.
- Connect QuickBooks read-only using the guides above. Ask questions of your real numbers instead of guessing.
- Skip the terminal unless you enjoy it. The gain for one set of books is smaller than it is for 30.
- Know when to hand off. A tool that answers questions about your books is not the same as books that are kept right every month.
Setup starting points for each tool
These are first steps, not full tutorials. Each one pairs a skill with a connection, because that pairing is the whole game.
Claude Code
- Get a paid Claude plan. Pro includes Claude Code.
- Make one folder per client. Add a CLAUDE.md with that client's notes.
- Write your first skill: a month-end checklist or a bank statement cleanup routine. See the Claude bookkeeping skill guide.
- Add the QuickBooks MCP server with one
claude mcp addcommand, per our Claude connection guide. Turn write access off. - Run the skill on one client. Read every change before you keep it.
Claude, formerly Cowork
- Use Claude on Pro or Max, where the new experience is rolling out first.
- Bring your skills and connectors. Reports say both carry over from Cowork.
- Keep Claude Desktop open for any task that needs local files or your browser.
- Set up the QuickBooks connection per our Claude connection guide, read-only.
- Hand off one long job. Review the output line by line before you use it.
ChatGPT desktop app with Codex mode
- Install the ChatGPT desktop app on macOS or Windows. ChatGPT plans include Codex, with limits that vary by plan.
- Pick a QuickBooks path from our ChatGPT connection guide. Start with the most limited access that does the job.
- In Codex mode, save your procedure as a reusable skill.
- Only schedule an Automation once the manual version has run clean a few times.
Codex CLI
- Install Codex CLI and sign in with ChatGPT (plan usage) or an API key (per-token billing).
- Package your procedure as a skill.
- Browse the plugin marketplace for the tools you need. Confirm a QuickBooks route before you build around one. The QuickBooks integrations hub is a good place to compare options.
- Run on exported CSVs first. Connect live books after the skill proves itself.
Frequently asked questions
These are the questions bookkeepers ask most about Claude Code vs Codex and the apps around them.
Is Claude Cowork gone?
Not really. Anthropic folded Cowork into Claude on September 16, 2026. The pricing page says "Claude Cowork is now just Claude," rolling out to Pro and Max first. Cowork-style tasks now run inside regular Claude conversations, and reports say connectors and skills carry over.
Is Codex free?
Codex is included with ChatGPT plans, including Free and Go, according to OpenAI. Usage limits vary by plan. Heavier use costs more: OpenAI's Pro plan starts at $100 a month. Codex CLI can also run on an API key, billed per token.
Do I need to code to use Claude Code or Codex CLI?
No. You need to be comfortable typing commands in a terminal and reading what the tool shows you. You're not writing software. You're pointing an agent at a folder and a written procedure. Expect a slow first week.
Can these tools change my QuickBooks books?
It depends on how you connect them. Some connection paths only read. Some can write. Our Claude and ChatGPT connection guides show what each path can do. Start read-only and open write access one client at a time, if ever.
Where Growthy fits
Every tool on this page helps you work on books. None of them is where the books live. A terminal agent with a good skill can clean a statement and draft categories. Then someone still has to review those categories and keep them consistent month after month.
That's the part Growthy handles. It categorizes transactions automatically. It's 85% accurate on first import, and you review the rest. You approve before anything counts. It works on top of QuickBooks Online, or from CSV imports, today. Use Claude Code or Codex for the one-off work. Use Growthy for the monthly work that has to be right every time.
If you're doing categorization by hand across a growing client list, try it on one client before your next close.
Growthy is bookkeeping software, not a CPA firm. This content is educational, not professional advice.
See It Work on Your Data
See Growthy on a sample book. Read-only bank access.
Bobby Huang • Partner, SDO CPA LLC / CEO, Growthy
Partner at SDO CPA. 18 years of hands-on bookkeeping. Bobby still reconciles real client books and builds Growthy from that operating work.
View author profileGrowthy content is written and reviewed by people who keep real books. Worked examples come from real bookkeeping scenarios, and product claims are checked against what the product does today. Our editorial guidelines cover how we source, verify, and update every article.
Keep reading

The Real Cost of Manual Bookkeeping: Time, Errors, and the Scaling Ceiling
You're good at this work. You know QuickBooks cold. You've developed a rhythm: open the client file, scan the transactions, start clicking through categories. Bank feeds, credit cards, merchant accounts. One by one.

Bookkeeping Automation in 2026: What Actually Works (and What's Just Marketing)
You've seen the demos. A tool ingests your bank feed, transactions appear pre-categorized, and the vendor calls it automated bookkeeping. It looks like magic until you're cleaning up 200 miscategorized transactions at month-end while your clients...

The Bookkeeper's Automation Stack: QBO + AI + Triage in 15 Minutes per Client
If you're managing 15+ QBO clients, you already know the math doesn't work.