Claude's finance plugins: the analyst research setup
Hey, it's Cindy 🌱 You commented ANALYST, so here is the whole setup from the reel. Anthropic open-sourced the plugin pack it built for financial services firms.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
Read this before anything else. Nothing here is financial advice, and I am not a financial adviser. These plugins draft analyst work product for a human to check: they do not make investment recommendations, execute trades, or tell you what to buy. Anthropic says the same thing in the repository itself ("Nothing in this repository constitutes investment, legal, tax, or accounting advice"). Every number a model gives you has to be checked against the primary source, which means the actual filing on the company's investor relations page or on SEC EDGAR, not a summary and not the model's memory. If a figure is not in the document you handed it, treat it as unverified.
what this actually is
The pack, in plain English
Anthropic publishes a public GitHub repository of reference agents and skills for finance work, the same ones it ships to banks and asset managers, and anyone can add it as a plugin marketplace (official tutorial). There are two layers in it, which is the bit most explainers get muddled:
🧩 Vertical plugins are bundles of skills and slash commands: financial-analysis (the core one, install first), equity-research, investment-banking, private-equity, wealth-management.
🤖 Agent plugins are self-contained workflow agents built on top of those skills: earnings-reviewer, model-builder, market-researcher and seven more.
The three worth your time as an individual investor are the earnings reviewer, the model builder and the market researcher. That is the trio in the reel.
Plan honesty. The plugins themselves are free and open source. They run inside Claude Cowork or Claude Code, and Cowork works with a paid Claude plan (claude.com/product/cowork). The repository also lists connectors to paid data providers like FactSet, Morningstar and PitchBook, and those genuinely need a subscription from the provider. You do not need any of them for this guide, because everything here runs on free public filings.
setup
Step 1: add the marketplace
Two routes, pick whichever matches how you use Claude.
🖥️ In Claude Cowork (the desktop app). Open the Claude desktop app, click the Cowork tab in the mode selector at the top, click Customize in the left sidebar, then Browse plugins, then Personal. Click the + button, choose Add marketplace from GitHub, and paste https://github.com/anthropics/financial-services-plugins.
⌨️ In Claude Code (the terminal). Run claude plugin marketplace add anthropics/financial-services. That registers the same catalogue under the name claude-for-financial-services.
🧱 Install the core plugin first. Whichever route you took, install financial analysis before anything else, because it carries the shared modelling skills the other plugins lean on. In Claude Code that is claude plugin install financial-analysis@claude-for-financial-services.
📦 Then add the three. Install earnings-reviewer, model-builder and market-researcher the same way. Once installed, the skills fire automatically when they are relevant, and you can also trigger them by typing / and picking one.
free inputs
Step 2: get the documents, without paying
The plugins are only as good as what you feed them. Handing Claude the real filing is what stops it answering from stale training data. All three of these are free.
SEC EDGAR The official US filings database. Full-text search every 10-K (annual), 10-Q (quarterly) and 8-K at sec.gov/edgar/search. This is the primary source, and it is the one to check figures against.
Investor relations pages Nearly every listed company posts its results deck, press release and webcast on its own IR page. Search the company name plus "investor relations". Often the clearest single summary of a quarter.
Earnings call transcripts The Motley Fool publishes free transcripts of most US earnings calls at fool.com/earnings-call-transcripts. This is what you paste in when you want the management commentary, not only the numbers.
Non-US companies Foreign companies listed in the US file a 20-F annual report instead of a 10-K, and it is on EDGAR too. Companies listed only overseas post the equivalent on their local exchange and their IR page.
plugin one
Earnings reviewer: read the call properly
This is the one that changes how a normal person reads results. It ships with /earnings-analysis, /earnings-preview, /model-update and /morning-note. You paste in the transcript or the filing, run the command, and it writes an analyst-style note: what management actually said, what changed against last quarter, and where the soft spots are.
Paste your transcript first, then this:
📞 the earnings read
I have pasted [COMPANY]'s [QUARTER] earnings call transcript and results release above. Work only from these documents and do not use anything from your training data for the numbers.
Write me an analyst-style note with these sections: what changed versus the prior quarter and versus guidance; the revenue and margin picture with the actual figures and the exact line or paragraph you took each one from; what management chose to emphasise and, just as importantly, what they were asked about and answered vaguely; any change in guidance or in the language around guidance; and the three things that would most change this picture next quarter.
Then do the part I care about most. State the bull case and the bear case in three sentences each, drawn only from this document, and tell me which specific number in the next report would settle the argument between them.
Finish with a list of anything I asked for that is not in the documents I gave you. Mark those as not found rather than filling them in, and do not give me a buy, hold or sell view.
Why the "not found" instruction matters. A model asked for ten things will usually produce ten things. Explicitly giving it permission to return a gap is what turns a confident-sounding summary into something you can actually rely on.
plugin two
Model builder: the numbers, in Excel
This one builds live spreadsheets rather than a wall of text. Its skills are /dcf-model, /3-statement-model, /comps-analysis, /lbo-model and /audit-xls. A discounted cash flow model is a structured way of asking what a company's future cash is worth today, and the value of building one is not the number it spits out, it is that every assumption behind the number is sitting in a cell you can argue with.
🧮 build a DCF you can argue with
Build a simple discounted cash flow model for [COMPANY] in Excel, using the filings I have given you for the historical figures.
Put every assumption on its own clearly labelled input tab: revenue growth per year, operating margin, tax rate, capital expenditure as a percentage of revenue, working capital, discount rate and terminal growth rate. No assumption should be hard-coded inside a formula anywhere in the model, because I want to change one cell and watch the whole thing move.
Show the historical figures next to the forecast so I can see whether the growth you have assumed is anything like what this company has actually delivered. Add a sensitivity table across discount rate and terminal growth so I can see how much of the valuation is coming from those two inputs alone.
Then write a short plain-English note under the model: which assumption the valuation is most sensitive to, which assumption you are least confident in, and where each historical number came from. Do not tell me whether the stock is cheap or expensive.
The trap with any DCF. Move the discount rate and the terminal growth rate a little and the valuation swings enormously, which means it is very easy to build a model that quietly confirms whatever you already believed. That is exactly why the prompt asks for the sensitivity table. Treat the output as a map of the assumptions, not a price target.
plugin three
Market researcher: the wider view
The third one zooms out from a single company to a sector or theme. Its skills are /sector-overview, /competitive-analysis, /comps-analysis and /idea-generation. This is the one to use before you get attached to a company, because it forces you to look at the competitors alongside it.
🗺️ the sector map
Give me an overview of the [SECTOR OR THEME] sector as it stands now, then place [COMPANY] inside it.
Cover: how the sector actually makes money and what drives demand; the five or six companies that matter most and roughly how big each one is; what structurally separates the leaders from everyone else; and the two or three forces most likely to reshape the sector over the next few years.
Then build a comparison table of [COMPANY] against its closest three to five competitors, with revenue growth, gross margin, operating margin and the common valuation multiples side by side. For every figure, give me the date and the source it came from, and mark anything you could not verify as unverified rather than estimating it.
Finish by telling me the single strongest argument that [COMPANY] is not the best business in this group. I want the counter-case in writing before I read anything else about it.
Turn on Claude's web search for this one, because sector context moves faster than any model's training data. Then click through the citations. Valuation multiples in particular change with the share price every single day, so treat any multiple as a prompt to go and check, not as a fact.
put it together
The full research pass, in order
1️⃣ Map the sector first. Run the market researcher prompt before you look closely at any one company. Deciding a company is great and then researching it is how confirmation bias gets in.
2️⃣ Read the last two calls. Pull two consecutive transcripts and run the earnings prompt on each. The interesting signal is almost always the change between them, not either one alone.
3️⃣ Build the model. Only now build the DCF, using the growth and margin figures the filings actually support rather than the ones that make the answer come out nicely.
4️⃣ Verify every number that matters. Take the five figures the conclusion rests on and find each one yourself in the filing on EDGAR or the IR page. If you cannot find it, it does not count.
5️⃣ Write your own view, in your own words. One paragraph, unaided, on what you think and what would change your mind. If you cannot write it without the AI, you have not finished the research.
the limits
What this setup cannot do
It cannot predict No model here forecasts a share price. It reads documents and structures analysis faster than you can, which is a different and much smaller claim.
It cannot verify itself A confident summary of a number it misread still sounds confident. The check against the primary filing is your job and it is not optional.
It has no live market data Not without a paid provider connector. Prices, multiples and anything intraday must come from a real source you have opened yourself.
It is not advice Anthropic designed these for qualified professionals to review before use. Same rule applies at home, with you as the reviewer, and consider a licensed adviser for anything that actually matters.