deskstack

Claude builds your DCFs and comps from SEC and European filings, and a second agent checks every number against
the filing. In the benchmark, 0 of 3,230 reported figures were wrong with deskstack, against 36 of 1,018 (3.5%)
with web search only.

If it saves you a tie-out, a star helps others find it.

Website · Examples · Tutorial · Tutoriel (FR)

Public beta. The hosted connector is online at https://deskstack.deskstack-server.workers.dev/mcp. Feedback
and issues are welcome.

Why

With web search only, Claude fills gaps with numbers that aren't in the filing. In the benchmark (50 companies, a
DCF and a comps table each):

• DCF: share counts derived from net income ÷ EPS, or taken from six months or one quarter; capex estimated as
  operating cash flow minus free cash flow; non-GAAP "adjusted revenue"; a segment's D&A used for the group's;
  figures from aggregator sites or Wikipedia. 20 wrong figures, in 12 of the 50 models. Claude flagged about half
  of them in its own notes, but the numbers went into the model.
• Comps: 15 of the 16 wrong figures were last-twelve-months figures a quarter late, from an aggregator that
  still showed the twelve months to June after the September 10-Qs were filed. 7 more were left as 0 placeholders.

deskstack gives Claude the filed figures, each with its filing, and reviewers that check every typed number. In
examples/apple-dcf-typo, Apple's FY2025 revenue is typed as 416,116: /associate says
the 10-K reports 416,161 and that two digits look swapped. /md-redteam shows what a price implies: at $230 a share,
the Apple example model implies a WACC of 7.74%, against the 10.6% it uses.

Benchmark

How many reported figures in an AI-built model are wrong, with and without deskstack. 50 companies (35 US, 15
European), three setups, Claude Sonnet 5.5. Every figure the scorer called wrong was read by hand against the filing.

Cost and time are averages per model. "Wrong" means the figure differs from the filing for the period asked. Not every
run finished: 49 of 50 DCFs with deskstack data, 45 of 50 comps tables with web search only. Analyses:
DCF, comps.

Read the 0 with care. It is partly by construction: the deskstack setups read the same data the scorer checks
against, so the 0 says the models copied deskstack's figures faithfully. Whether those figures are right is checked
separately, by the golden set and a nightly re-read with other XBRL parsers (Accuracy). Each analysis
also lists what the run found wrong in deskstack itself, and what was fixed. With 30 searches per request and high
effort instead of 12 and medium, web search still wrote a wrong figure (1 of 117), and left none as 0
(robustness check). The runner, the frozen company set and the method are in
bench/; the chart version is bench/benchmark.svg.

Set up in the Claude app

deskstack has two parts. The data connector gives Claude filed figures, each with its source; you add it once,
by pasting a URL. The plugin adds the roles: commands you type in a chat, such as /analyst or /associate.

The connector works on any Claude account; a free account can add one custom connector. The plugin needs a paid
plan, and code execution turned on in Settings → Capabilities. No terminal is needed.

1. Data: in Claude, open Customize → Connectors → Add custom connector and paste
   https://deskstack.deskstack-server.workers.dev/mcp, with no sign-in. Check it: ask "Using deskstack, what was
   Apple's revenue in FY2025? Cite the filing." The answer should be $416,161 million, citing an Apple 10-K and its
   accession number. You can also run your own copy (self-hosting).
2. Roles: open Customize → Plugins → Add marketplace, paste https://github.com/joaquimmj/deskstack, then
   install deskstack.
3. Try it: /deal-team Apple quick-dcf. If something fails, /deskstack-doctor checks the connector, the plugin
   version, Python and the bundled scripts, and says how to fix what doesn't work.

The tutorial walks through a first reviewed DCF in about 10 minutes: English ·
français. In Claude Code: /plugin install deskstack --marketplace joaquimmj/deskstack (2.1.275+).

What you get

A workbook with blue inputs, formulas, and a Sources sheet that cites the filing behind every typed number:

Review reports from each role, and one sign-off page. Nothing is exported until you approve it in your own words:

Both images are rendered from the files in examples/apple-dcf by
docs/media/make_media.py. Open the workbook and every report there, or the
four other examples, including a model with errors typed in on purpose.

What it catches

Real errors from the benchmark, each read by hand against the filing.

To see these caught, open examples/errors-caught: a Microsoft DCF with four of these errors
typed in on purpose, and the reports that find all four.

verify_value also names the usual causes of a mismatch: wrong scale, opposite sign, another period, another line
item, another currency, two digits swapped, or another common definition (net debt with leases, equity without
minorities). The 21 model rules are listed in checker/README.md.

Check any AI-built model in one command

For teams with a terminal or a repository; the Claude app doesn't need it.

deskstack-check checks any .xlsx model, whether Claude, ChatGPT, Copilot, another tool or a person built it.

• Offline: the 21 model checks (hardcodes, broken rows, balance sheet, terminal value, placeholders…).
• With --verify: every typed figure is also tied to the SEC or ESEF filing through the deskstack connector.
  Only those figures are sent, each with its line item, period and cell reference (e.g. Inputs!C5).

pipx run deskstack-check model.xlsx            # soon: the package isn't on PyPI yet
pipx run --spec ./checker deskstack-check model.xlsx   # today, from a clone

$ deskstack-check examples/apple-dcf-typo/model.xlsx --verify
examples/apple-dcf-typo/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)

Options:

• --markdown or --json: other output formats.
• --company AAPL: the company, when the workbook doesn't name it as "Name (TICKER)".
• --connector URL: your self-hosted server.
• --ignore DS010,DS017: leave out rules your team accepts on purpose; the report says which.
• --fail-on major: the level that fails. The default is blocker; a figure that doesn't tie counts as one.

The exit code is 0 for a pass, 1 for a fail, and 2 when the file can't be read or the connector fails.

In a repository, 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"        # optional: tie figures to the filings
    fail-on: blocker

The roles

flowchart LR
  A["/analyst<br>builds the workbook<br>from cited data"] --> B["/associate<br>ties every number<br>to the filing"]
  B --> C["/vp-review<br>21 model checks"]
  C -- "findings you accept" --> F["/analyst<br>fixes, rebuilds"]
  F -- "reviews run again" --> B
  C --> D["/md-redteam<br>sensitivities,<br>objections"]
  D --> E["/compliance<br>disclaimer, citations"]
  E --> G["/sign-off<br>one page of every check"]
  G -- "your approval,<br>in your own words" --> H[("signed copy")]

Reviewers report; they never edit. In Claude Code and Cowork, they run as separate agents that don't see the
builder's reasoning. Details: docs/ROLES.md.

Works with Anthropic's financial-analysis plugin

Built your model with Anthropic's financial-analysis plugin? Run /associate, /vp-review, /md-redteam and
/sign-off on that workbook as it is. /vp-review merges the plugin's audit-xls findings with its own and marks
the ones both raised. Details: docs/ROLES.md.

Data sources and coverage

Every value carries its source: filing, accession or report, period. The 8 tools are documented in
docs/TOOLS.md. Terms and attribution for each source are in docs/DATA_SOURCES.md.

Accuracy

• The golden set: 377 questions with known answers from the filings. It covers:
  • US issuers and IFRS 20-F filers;
  • at least 10 questions for each of the 14 European countries covered, the UK included, each answer read by
    Arelle from the published report;
  • 47 questions for Brazil, 10 of them on the last twelve months, each answer read from CVM's published CSVs by a separate Python reader
    (evals/cvm_reader.py), since no second parser exists for them;
  • a bank;
  • restatements and a stock split;
  • fallbacks that must say what they are (D&A that is depreciation only, debt with finance leases, equity with
    minorities, net debt without leases), with the caveat carried into the ratios;
  • 52/53-week years, and fiscal years ending in January, February, March, June and September.
• Every push: every question must pass through the real tools on recorded data. Today: 100%.
• Nightly: an independent re-read of each answer from the filing itself, with edgartools for SEC and Arelle for
  ESEF. Details: evals/README.md.

Limitations

• No share prices: a market cap needs a price you give.
• No consensus estimates.
• European and UK coverage is partial and annual only. There are no German issuers, because their reports aren't in
  the public ESEF repository. Some UK companies tag no total assets, and some issuers tag only parts of revenue (goods and services, as NOS,
  REN, Sonae and Altri do); those lines are then null, with a note.
• Brazilian coverage is 20 companies, annual years plus LTM, from CVM's fixed chart of accounts. Capex, D&A and EPS aren't
  fixed lines there, so they are null, and so is EBITDA. Total equity includes minority interests; the minorities
  are served on their own line and the equity note gives the amount. Lease liabilities aren't a fixed line either.
  Banks and insurers aren't covered yet.
• IFRS 20-F filers have no LTM.
• When a filer tags a line only in a broader or narrower form (D&A without amortization, capex with intangibles,
  debt with finance leases, interest as total finance costs), the figure is served as filed with a note, not
  adjusted. Ratios built on it repeat the note.
• Banks and insurers are flagged; EBITDA and EV metrics don't apply to them.
• The benchmark's 0 is partly by construction (see Benchmark). It ran on one model, Claude Sonnet 5.5.
• It checks figures against filings. It never says buy or sell, and it is not investment advice.
• Reviewers reduce errors but don't replace a human reviewer. Everything is a draft until you sign it off.

How it stays free

• Free hosting: the server runs on Cloudflare Workers' Free plan, with a daily cap and a per-IP limit.
• Static snapshot: SEC data comes from a nightly snapshot served as static files.
• Self-hosting: your own copy costs nothing. scripts/self-host.sh deploys it in one command
  (docs/SELF_HOSTING.md).
• When the shared server is busy: past the daily cap or the per-minute limit, every tool says what happened, when
  service comes back (the daily cap resets at 00:00 UTC) and how to self-host.
• Privacy: the server stores no user data and logs no prompts (PRIVACY.md).

How it compares

FAQ

Does it work on the free plan? The data connector does: a free account can add one custom connector. The roles
are a plugin, and plugins need a paid plan.

Claude Code or Cowork? Both work, with the same plugin. There, the reviewers run as separate agents that don't see
the builder's reasoning (docs/ROLES.md).

Which companies are covered? US-listed companies that file XBRL with the SEC, including IFRS foreign filers; 232
large European and UK issuers; 20 Brazilian companies. company_lookup says when one isn't covered
(coverage).

Is my data stored? No. The workbook checks run where Claude runs your skills: on your machine in Claude Code, in
Claude's sandbox in the apps. Only the tool calls reach the deskstack server, which stores no user data and logs no
prompts (PRIVACY.md).

Why no share prices? We found no free price source whose terms allow serving prices to everyone
(docs/DATA_SOURCES.md). A market cap uses a price you give, with its date.

Can I self-host? Yes, at no cost: scripts/self-host.sh deploys your own copy on Cloudflare's free plan
(docs/SELF_HOSTING.md).

Contributing

You can add a data source, a checker rule, a role, golden questions or benchmark companies. See
CONTRIBUTING.md.

cd server && npm ci && npm test           # data server (TypeScript, Cloudflare Worker)
python3 -m pytest checker/tests           # checker and plugin scripts (Python, openpyxl)

Credits

• SEC XBRL standardization: first solved by open-finance-mcp
  by Nathanael Johnson (MIT). deskstack ports it to TypeScript with parity tests.
• Bundled code: openpyxl and et_xmlfile (MIT), vendored in the plugin.
• Third-party data: SEC EDGAR, XBRL International's filings.xbrl.org, U.S. Treasury, ECB statistics, and Aswath
  Damodaran (NYU Stern).
• Brazilian data: contains information from the CVM Open Data Portal (https://dados.cvm.gov.br/), made available
  under the Open Database License (ODbL). The Brazilian data deskstack derives is published under the ODbL too, and
  is free to download whole (v1/br/dataset.json on the server). The code stays MIT.

See THIRD_PARTY_NOTICES.md.

Independent open-source project. Not affiliated with or endorsed by Anthropic. Data may be incomplete or delayed;
verify against the cited filing. Not investment advice. MIT license.

Built by the team behind Aptence.