Skip to content
CodelessOps
Go back

What Anthropic's finance plugin leaves to you

Karim Lameer

Karim Lameer — Master Anaplanner, CIMA-qualified, 15 years in finance and FP&A. I build Grounded and published the Board Pack Test. LinkedIn · New here? Start here

I expected Anthropic’s reconciliation skill to get the numbers wrong. So I gave it a month of a fictional distributor’s bank statement, cashbook and last month’s reconciliation, and ran it three times, each in a fresh folder. It was right to the penny all three times. It even found the cheque nobody had chased since April.

Then I read how it got there. Each run wrote its own code to do the matching, and the code was different each time. Each run ended with the sign-off line left blank for me to fill in. One called a 47-day-old cheque “stale”, a word the skill’s own aging table keeps for items over 90 days.

That gap is what this piece is about: what is in Anthropic’s finance plugin, what it does well, the three things a month-end close needs that it leaves to you, and where to get those three free.

A plugin is a folder of instructions

A plugin is a folder you add to Claude, in the desktop app, in Cowork or in Claude Code. Anthropic’s finance plugin is free and open source, designed for Cowork and usable in Claude Code too.1 You install it in two lines and Claude gains a set of finance skills.

Each skill is a page of written instructions for one task. There are eight: reconciliation, variance analysis, close management, journal entries, journal entry preparation, financial statements, audit support and SOX testing.2 When you ask for a bank reconciliation, Claude reads the reconciliation page and follows it. Alongside the skills sit connector placeholders, such as one marked for your ERP. No ERP connector ships, so you bring the data yourself.3

What isn’t in the folder matters more: there are no programs. Every skill is prose, and where a figure is needed, the page tells Claude to work it out, as in “For each line item, calculate and flag material variances.”4

Anthropic's finance plugin, opened finance/ manifest (name, version) 8 skills written instructions, read by Claude, one per task connector placeholders ERP: none ships, you bring data programs none So when a figure is needed, Claude works it out itself, in whatever way it chooses that day.
Everything in the folder is text. The arithmetic is left to the session that reads it.

A second Anthropic repository packages the close as named “agents”: a Month-End Closer, a GL Reconciler and a Statement Auditor. That design does enforce some things properly. The part that reads untrusted invoices cannot write anything, and only one part writes the output.5 The GL Reconciler adds a critic, a second model that re-checks every break it reports.6 Approval, though, is a sentence: posting “requires controller approval outside the agent.”7 Keep that in mind, along with the missing programs.

The method in it is good

I’d have written much of this method myself, and in my own close kit I did. The reconciliation skill sorts reconciling items into three kinds: timing differences that clear on their own, adjustments that need a journal, and items to investigate, with an action for each.8 Its escalation table says an unreconciled difference of any amount means you cannot close until it is resolved or documented, and its best practices say “do not just carry items forward indefinitely”.9

The variance skill insists that commentary names the actual driver, not just “higher than expected”.10 The fund-administration skills go further. A roll-forward must foot, and if it doesn’t, “surface it, don’t plug it”.11 Where the data doesn’t show a cause, the commentary says “driver unclear” and goes to the controller instead of inventing one.12 My own kit has the same rule in nearly the same words: no cause you cannot source.

Anthropic is clear about what these are. Its finance agents are “starting points”, to install and then tune to how your firm works.13 Its own blog carries a customer’s best practice that “Computation is a job for a formula engine”, and the same repository gives its pitch-work agent Python for exactly that.14 15 This is a good checklist. A close needs more than a checklist.

Three runs: right every time, different code every time

I took the reconciliation skill exactly as Anthropic publishes it and checked every file against the repository. I gave it the same month of data three times, each in a fresh folder with my own answer key removed, and asked the same question each time. The data is fictional and clean: one bank account, one month, an easy match. Three runs is a demonstration, not a statistic.16

All three runs reconciled exactly: 9,882,613.23 on the statement, 8,911,388.92 in the books, six reconciling items between them, difference nil. All three found cheque 004182 for 1,850.00, written in April and still not cashed. Each took about a minute.

Each run wrote its own Python to match the lines and foot the balances, which is what I’d have asked for. But the code was different every time (two runs took the same approach, none wrote the same program) and none of it was kept. Next month there would be a fourth version.

Each run also ended by handing the sign-off back to me: “left blank for you to complete” in one, “a reviewer needs to sign off” in another. All three noticed that the 1.19m month-end payment run was above the skill’s own example threshold for controller review, and finished anyway, because the threshold is a line in a table.17 And the old cheque came back three ways: in the 31 to 60 day “Aging” band, as “stale”, and as “Timing, but aging”. The skill’s own table keeps “stale” for 90 days and over.

Same data, same prompt, three runs Run 1Run 2Run 3 Bank to cashbookReconciling itemsOld cheque foundCode it wroteThe old cheque, calledSign-off agrees, 0.00agrees, 0.00agrees, 0.00 666 yesyesyes 107 lines97 lines91 lines b84f7a0ed318f490d5ee6c46 "Aging", 31 to 60"stale""Timing, but aging" blankblank"needs to sign off"
The answer was the same every time. The code, the judgement words and the empty sign-off were not under anyone's control. The fingerprints are the first characters of a hash of each run's code.

To be fair to the test, a single run like this can’t pause halfway for a signature. The point is not that it carried on. It’s that nothing in the plugin would make the next step wait for one. Right three times, and the sign-off left blank three times.

Right isn’t enough: one tested program, a reviewer who must sign, a sealed record

Anyone who has to rely on your close will ask three questions about a number in it. Your auditor will, your finance director will, and so will whoever has your job after you. How did you get it? Who checked it? Can you show me?

How did you get it? The answer a close needs is one tested program: the same code every month, checked against answers you already know. Auditors test a figure by recalculating it and re-performing the control that produced it.18 That works when the code is fixed and the inputs are kept (I went through what this means for a finance-built close in how to build it). A program improvised for one run, and gone by the next, can be right but cannot be re-performed.

Who checked it? The answer is a reviewer who must sign: a step that waits for a named person before the next one starts. In the plugin, review is a sentence at the top of each skill asking that outputs be reviewed by a qualified professional. Nothing makes the variance report wait until the bank reconciliation is signed. My own rule for an item like that old cheque is simple. Past the limit you set, it stops being a timing difference and becomes a decision, reissue, write back or carry forward, and a decision belongs to a person with a name.

Can you show me? The answer is a sealed record that ties each figure to the person who signed it. Anthropic’s managed agents do keep “a full audit log” of every tool call,19 and Cowork sessions reach the enterprise compliance API.20 Those logs show what the model did. They don’t show that a named person signed a specific number, or that the number in the pack is the number they signed.

What the person relying on your close asks, and what answers it How did you get it? Who checked it? Can you show me? One tested programsame code every month, checked against known answers A reviewer who must signthe next step waits for a named person A sealed recordwho signed which figure, and proof it hasn't changed Underneath all three: the method. That is the part Anthropic's plugin supplies.
The plugin answers "what should be done". A close also has to answer how, who, and where's the proof.

The obvious reply is to tell Claude to save the script. It’s a good first step, and it isn’t enough on its own. A saved script still needs tests that show it is right on next month’s data, a step that won’t run until someone signs, and a record of who signed what. That is the difference between a prompt and a control. It is also how a close I rebuilt went from twelve working days to under eight hours: the logic moved out of linked workbooks into one governed place, so it ran the same way every month and kept running after I left.21

None of this is a new demand invented for AI. The UK’s audit regulator says language models are “often not natively strong at computational tasks” and should be given tools for them, with human review and a record of intermediate outputs.22 US audit staff report preparers worrying that inconsistent AI output raises questions about auditability.23 Neither prescribes a mechanism. They ask for what any careful reviewer would.

All three fit in the same folder, free

You don’t need a new platform to add these. They fit inside the same plugin format, and I’ve built one way to do it, called verified-close.24 It keeps Anthropic’s eight skills word for word, gives four of them programs, gates and a close log, and marks every file it changed, as the licence asks.

For the first, scripts compute every figure in the bank reconciliation, the variance report and the management pack. Each figure is re-derived by independent checks before the reviewer sees it, and again at sign-off. A small golden set of known answers runs after any change, so a broken calculation turns it red.

For the second, each step’s script refuses to start until the step before it is signed in the close log. An old cheque can’t be signed without a recorded decision on it.

For the third, every sign-off seals the signed files with a fingerprint (a sha256 hash) in the close log. If anyone edits a signed workbook afterwards, the next step stops. At close-out the log becomes an audit binder: who signed each step, when, what they decided, and the fingerprint of every file.

The gated close: compute, stop, sign, seal 1. Bank reconciliationscript matches, checks to the penny 2. Variance reportscript computes, Claude writes notes 3. Management packbuilt from signed figures only Reviewer signsand decides the old cheque Reviewer signs Reviewer signs fingerprintfingerprintfingerprint Close log who signed, when, what they decided, the hash of every signed file. At close-out it becomes the audit binder.
Each step's script won't start until the step before it is signed, and a signed file that changes stops the close.

I put the gates in the scripts rather than in the instructions, because an instruction can be talked round when the deadline is close. In the recorded run I asked Claude to skip the sign-off because I was short of time. It refused, and if it hadn’t, the next script would have refused anyway. The run, its transcript and every filed document are in the repository, and the eight sealed files check out against the log.24

It runs on a fictional company with one bank account, and it’s built to be read. Your ledger, your chart of accounts, your materiality and your reviewers are the part that changes for a real close.

Read the run, take the repo

If you want to try Anthropic’s plugin on Monday, run it on a month you have already closed and signed, so you know the answers before it gives you its own. Then ask it the three questions. How did you get this? Who checked it? Can you show me?

Read the recorded run first, then take the repository. It’s free, and Anthropic’s skills are inside it unchanged.24 If you want it running on your own ledger, that is the work I do. Next in this series: how I test an agent skill before I trust it with a close.

Sources

Footnotes

  1. Anthropic, finance plugin README, knowledge-work-plugins at commit 33eeda1, October 2026. Designed for Cowork, also works in Claude Code. ↩

  2. Anthropic, finance plugin folder, knowledge-work-plugins at commit 33eeda1. Thirteen files: eight skills, manifests, connectors, licence. No scripts, hooks, commands or agents. ↩

  3. Anthropic, CONNECTORS.md, at commit 33eeda1. ERP and accounting: “no supported MCP servers yet”. ↩

  4. Anthropic, financial-statements skill, line 99, at commit 33eeda1. ↩

  5. Anthropic, Month-End Closer managed-agent template, poster subagent, financial-services at commit 574ed36. The invoice reader gets read-only tools; only the poster can write. For builders: the poster’s skill tells it to run Python with Bash, but its tool list has no Bash, and the deploy script’s documented schema-validation wrapper is not wired in (deploy-managed-agent.sh, lines 9 to 10). The close agents’ own tool lists hold no shell (month-end-closer.md, line 4). ↩

  6. Anthropic, GL Reconciler critic subagent, at commit 574ed36. ↩

  7. Anthropic, Month-End Closer agent, line 28, at commit 574ed36. ↩

  8. Anthropic, reconciliation skill, lines 90 to 122, at commit 33eeda1. ↩

  9. Anthropic, reconciliation skill, lines 161 to 173, at commit 33eeda1. ↩

  10. Anthropic, variance-analysis skill, lines 145 to 159, at commit 33eeda1. ↩

  11. Anthropic, fund-admin roll-forward skill, line 29, at commit 574ed36. ↩

  12. Anthropic, fund-admin variance-commentary skill, line 30, at commit 574ed36. ↩

  13. Anthropic, financial-services README, line 17, at commit 574ed36: “They’re starting points: install the ones that match your work, then tune the prompts, skills, and connectors to how your firm does it.” ↩

  14. Anthropic, How Kepler built verifiable AI for financial services with Claude, claude.com blog, April 2026. A customer case study; the quoted line is from Kepler’s list of best practices. ↩

  15. Anthropic, pitch-agent modeler subagent, line 6, at commit 574ed36: “Run calculations in Python via Bash”. ↩

  16. The three runs, with every line of code Claude wrote and its full answers, in the verified-close repository. Plugin files verified byte for byte against knowledge-work-plugins at commit 33eeda1; Claude Code run headless with no other plugins, settings or connectors; 18 October 2026. ↩

  17. Anthropic, reconciliation skill, aging table line 133 and escalation table lines 156 to 158, at commit 33eeda1. “90+ days: Stale”; individual item over 50,000: controller review. ↩

  18. BDO, Understanding and managing information produced by the entity in audits, August 2024. Accuracy is tested by tracing data to the system and recalculating to replicate results. ↩

  19. Anthropic, Claude for financial services: finance agents, May 2026. ↩

  20. Anthropic, Cowork for Team and Enterprise plans, Claude Help Center. Cowork sessions are captured by the Compliance API for the most recent 30 days. ↩

  21. Karim Lameer, 12 days to 8 hours: what actually moved the close, codelessops.com, September 2026. ↩

  22. Financial Reporting Council, Generative and Agentic AI Guidance, March 2026. ↩

  23. PCAOB staff, Spotlight: Observations on the use of generative AI, July 2024. ↩

  24. Karim Lameer, verified-close, Apache-2.0, October 2026. Built on Anthropic’s finance plugin; every changed file is marked and the upstream commit is pinned. ↩ ↩2 ↩3


Want the manual work around your close taken out?See how I build it: scoped first, built in phases at a fixed price, with a named person approving every output.

Share this post:

Keep reading

All posts →