A deal team inside Claude. One agent builds the model from free, cited data. Others tie every number to the SEC or European filing. Nothing ships until you sign off.
50 companies × DCF and comps × 3 setups, Claude Sonnet 5.5, November 2026. The zero is partly by construction: the scorer checks against the same data, which is checked every night against two other parsers.
Each role is a skill with its own checklist and a fixed report. Reviewers report; they never edit your file.
History, from the filings
| FY2023 | FY2024 | FY2025 | |
|---|---|---|---|
| Revenue | 383,285 | 391,035 | 416,161 |
| EBIT | 114,301 | 123,216 | 133,050 |
| D&A | 11,519 | 11,445 | 11,698 |
| Capex | 10,959 | 9,447 | 12,715 |
Source: Apple 10-K FY2025 · 0000320193-25-000079 · sec.gov
Inputs
| WACC | 9.0% |
|---|---|
| g | 3.0% |
| Tax | 16.0% |
13 of 14 figures match Apple's 10-K. 1 mismatch.
| Line | Model | Filed | |
|---|---|---|---|
| Revenue FY2025 | 416,161 | 416,161 | |
| Revenue FY2024 | 391,053 | 391,035 | |
| EBIT FY2025 | 133,050 | 133,050 | |
| D&A FY2025 | 11,698 | 11,698 | |
| Capex FY2025 | 12,715 | 12,715 |
Typo: two adjacent digits look swapped (filed 391,035, written 391,053).
Model check: clean0 blocker · 0 major · 0 minor
Compliance: 10 pass, 0 fail
| Tie-out to filings | pass | 14 match |
|---|---|---|
| Model review | pass | 0 blocker, 0 major |
| Compliance | pass | 10 pass, 0 fail |
Approval token 3ba96d58cd17
Nothing is exported until you approve in your own words.
In Claude Code and Cowork the reviewers run as separate agents that never see how the model was built. They review work you bring, too: any .xlsx.
verify_value reads the filed figure for the same company, line and period, and says why a number differs: scale, sign, currency, period, line item or a typo.
Every mismatch comes back with the filed value, the filing's accession number and the likely cause.
The review scripts on Apple's FY2025 DCF with one planted typo. The model review can't see it; the tie-out to the 10-K can.
Claude Sonnet 5.5 built a DCF and a comps table for each of 50 companies (35 US, 15 European), three ways. Every reported figure was scored against the filing.
| Setup | Reported figures wrong | Cost / model | Time |
|---|---|---|---|
| Web search only | 3.5% 20 of 573 | $0.23 | 58 s |
| deskstack data | 0% 0 of 593 | $0.07 | 18 s |
| deskstack + reviewers | 0% 0 of 606 | $0.10 | 32 s |
| Setup | Reported figures wrong | Cost / model | Time |
|---|---|---|---|
| Web search only | 3.6% 16 of 445, and 7 left as 0 | $0.56 | 171 s |
| deskstack data | 0% 0 of 1,021 | $0.17 | 36 s |
| deskstack + reviewers | 0% 0 of 1,010 | $0.22 | 51 s |
In the DCFs, numbers it couldn't find and filled in anyway: derived share counts, estimated capex, non-GAAP revenue. In the comps, twelve-month figures a quarter out of date, from an aggregator.
A pre-registered re-run with 30 searches and high effort still wrote a figure that isn't the filing's. The placeholders went away; the cost per table more than doubled.
Robustness checkThe runs found gaps in deskstack's own data and two bugs in the scorer. All fixed, tested, and written up with the results.
Claude with web search alone made these errors building DCFs and comps tables. Each was read by hand against the filing, and each is caught by a deskstack check.
| Error | Real example | What catches it |
|---|---|---|
| Last twelve months a quarter late | FedEx, Fortinet, Motorola Solutions: an aggregator's "TTM" to June, after the September 10-Qs (15 figures) | /associate: verify_value says Stale and gives the latest twelve months |
| A quarter written as the year | Boeing net income: the fourth quarter (8,220) as the full year (2,235) | /associate: mismatch against the filed annual figure |
| Share count derived or from the wrong period | Boeing, Pfizer, GE diluted shares: net income ÷ EPS, a six-month count, a Q1 count | /associate: mismatch, and a flag when diluted shares are more than 10% from the cover page |
| Non-GAAP or segment figure for the group | GE "adjusted revenue"; Stellantis D&A for industrial activities only | /associate: mismatch against the group's filed line |
| Figure from an aggregator or Wikipedia | Cisco and Verizon operating income; Schneider operating income, rounded | /associate: mismatch, with the filed value and its source |
| Estimate typed as a filed figure | Netflix capex as operating cash flow minus free cash flow | /associate: mismatch; rule DS016 also flags a figure labelled as an estimate with no filing cited |
| Placeholder left at 0 | Cigna, CVS and Humana in UnitedHealth's comps, labelled "NOT FOUND" (7 figures) | /vp-review rule DS016: a typed 0 labelled as not found |
deskstack-check checks any .xlsx model, whether Claude, ChatGPT, Copilot, another tool or a person built it. Offline, it runs the 21 model checks. With --verify, it also ties every typed figure to the filing; only those figures are sent, each with its line item, period and cell reference.
pipx run deskstack-check model.xlsx --verify$ deskstack-check model.xlsx --verify
model.xlsx (deskstack_check 0.2.0, sha256 ced59305c2dd)
Model checks: 0 blocker, 0 major, 0 minor, 0 info
Tie-out to the filings (AAPL): 13 of 14 match, 1 do not
MISMATCH Inputs!C5 Revenue FY2025 416,116 filed 416,161
Typo: two adjacent digits look swapped (filed 416,161, written 416,116).
FAIL (fails on blocker)Soon on PyPI. Today, from a clone: pipx run --spec ./checker deskstack-check model.xlsx.
The GitHub Action checks every workbook on each push and writes the report in the job summary.
- uses: joaquimmj/deskstack@main
with:
files: "models/**/*.xlsx"
verify: "true"
fail-on: blockerBuilt with deskstack's real tools from recorded responses: the workbook, the tie-out, the model review, compliance and the sign-off page.
14 of 14 filed figures tie to the 10-K. Ready for sign-off.
examples/apple-dcfApple against Microsoft, Alphabet, Amazon, Meta and Nvidia: 24 of 24 figures tie. Market caps stay empty without a price.
examples/us-tech-compsFrom its ESEF annual reports, in euros. 14 of 14 tie; the review flags a terminal value at 82% of EV.
examples/heineken-dcfA Microsoft DCF with four benchmark errors typed in on purpose. All four are caught.
examples/errors-caughtThe leaderboard compares models on the same 50 companies: Claude Sonnet 5.5 so far, more soon.
Official filings and public datasets only. A figure that isn't filed stays empty; it is never estimated.
10-K, 10-Q and 20-F filings: annual and last twelve months, for US-listed companies including foreign filers under IFRS.
ESEF and UKSEF annual reports of 232 large issuers in 14 countries: 72 UK listed companies, and issuers from France and the Netherlands to Spain, Italy and the Nordics.
Annual DFP and quarterly ITR statements of 20 large listed companies, by CVM id or B3 ticker. Under the ODbL: the derived data is free to download whole.
Daily yields from 1 month to 30 years.
30+ currencies, on a date or averaged over a period.
Equity risk premium by country; betas and cost of capital by industry.
Source, URL, accession or ID, period and as-of date on each figure, so you can open the filing it came from.
Golden questions answered exactly as filed, against the live connector, and every answer re-read from the filing by an independent parser.
You need a Claude account with code execution on; the roles need a paid plan.
In Claude, open Settings, then Connectors, then Add custom connector. Paste the URL, with no sign-in.
Coming soonThe hosted connector isn't online yet. You can run your own copy for free in a few minutes: self-hosting guide.
Open Customize, then Plugins, then Add marketplace. Paste this repository, then install deskstack.
https://github.com/joaquimmj/deskstackClaude builds the DCF, has every number checked and reviewed, then waits for your approval.
/deal-team Apple quick-dcfNo account, no card, no trial. The hosted connector has a daily capacity shared by everyone; for more, run your own copy.
Paste one URL into Claude and start.
$0 forever
Install in ClaudeYour own copy, on your own free Cloudflare account.
$0 on Cloudflare's free plan
Self-hosting guideYes. The data connector runs on Cloudflare's free plan with a daily cap shared by everyone, and the code is MIT-licensed. For heavy use, deploy your own copy in a few minutes: self-hosting guide.
From the filings themselves: SEC EDGAR for US companies, ESEF reports on filings.xbrl.org for European ones, plus the U.S. Treasury, the ECB and Damodaran's datasets. Every figure carries its source.
No. The review scripts run in your own Claude session. The server only sees the values the associate sends to verify_value, and keeps no user data. Privacy.
No share prices and no consensus estimates: a market cap needs a price you provide. European coverage is partial and annual only. Banks and insurers are flagged, since a standard DCF doesn't fit them.
No. It catches what can be checked: figures against filings, formulas against rules, language against a checklist. Everything is a draft until you sign it off.
No. deskstack is an independent open-source project that works with Claude. It is not affiliated with or endorsed by Anthropic.