finance-agent-kit

Equity valuation skills and data MCPs for DeepSeek Harness, Claude Code, and any harness that reads the SKILL.md convention.

Seven valuation skills from Anthropic's Apache 2.0 financial-services repository, plus two MCP servers that supply the data those skills assume you already have from a Bloomberg terminal.

Install from an agent

If an agent is doing this for you, these are the steps. Run them from the project directory, and report back rather than mounting anything.

1. Confirm the project has a .git directory. If not, run git init first: dsh resolves the project root by walking up for one, and without it the skills load from the wrong place or not at all.

2. Download bootstrap.py from https://raw.githubusercontent.com/davidalmeida90/finance-agent-kit/main/bootstrap.py and read it before running it.

3. Check these import, and install any that do not: mcp, yfinance, pandas, lxml, edgartools, openpyxl, python-docx, matplotlib. On Windows, pywin32 as well.

4. Run bootstrap.py, passing --identity "<name> <email>". That value is the SEC's fair-access requirement, so ask the user for it rather than inventing one.

5. Report the skills now visible and the two patch files written. Do not start the harness. The composition is read at boot, so mounting is the user's step:

   dsh --profile web --patch ./sec-edgar.cordis.yml --patch ./market.cordis.yml --port 3081
   

   Port 3081 rather than the 3080 default, because the session you are running in is almost certainly already serving on 3080. Only one process can hold a port, so reusing it fails with EADDRINUSE and the new harness never starts. Drop --port only if nothing is serving on 3080.

Installing writes an AGENTS.md into the project carrying the data, cost of capital and output rules. Any existing one is kept as AGENTS.md.previous. Pass --no-agents-md to keep yours instead.

Install

From inside your project directory. No git needed, nothing left over.

curl -sL https://raw.githubusercontent.com/davidalmeida90/finance-agent-kit/main/bootstrap.py -o bootstrap.py
py -3 -m pip install mcp yfinance pandas lxml edgartools openpyxl python-docx matplotlib
py -3 bootstrap.py --identity "Your Name you@example.com"

On PowerShell, swap the first line for iwr https://raw.githubusercontent.com/davidalmeida90/finance-agent-kit/main/bootstrap.py -OutFile bootstrap.py.

Read bootstrap.py before running it. It is 100 lines and it downloads code from the internet, which is the category of script worth reading first.

Then start the harness:

dsh --profile web --patch ./sec-edgar.cordis.yml --patch ./market.cordis.yml --port 3081

You end up with the skills in .dsh/skills, two resolved patch files, and a finance-agent-kit/ directory holding the MCP servers. Keep that directory: the patches point at it.

Or clone, if you prefer

git clone https://github.com/davidalmeida90/finance-agent-kit.git
cd finance-agent-kit
py -3 -m pip install -r requirements.txt
py -3 install.py --target /path/to/your/project --identity "Your Name you@example.com"

Two things that catch people

--identity is an SEC fair-access requirement, not a secret or a credential. It is the name and email the SEC sees on your requests.

Your project needs a .git directory. dsh finds the project root by walking up looking for one, so without it your skills resolve to some parent directory and quietly fail to load. Run git init first.

Add --skills-dir .agents/skills if you want the same install to work in Claude Code as well as dsh.

What is in it

Skills

Skill	Does
dcf-model	DCF with scenarios, WACC build, sensitivity tables, Excel output
comps-analysis	Trading comparables with statistical benchmarking
3-statement-model	Linked income statement, balance sheet, cash flow
audit-xls	Formula tracing and workbook audit
xlsx-author	Spreadsheet construction conventions
initiating-coverage	Five task initiation pipeline ending in a DOCX report
earnings-analysis	Post-print earnings notes

Vendored unmodified. Provenance, commit hash and known defects in skills/VENDORED.md.

MCP servers

sec-edgar wraps EdgarTools. Keyless, read only, and the authority for anything an issuer reports. Company financials, individual filings, narrative sections, and the notes where segment and geographic revenue actually live.

market-data is original, and exists because filings do not carry share prices and nothing assembles trailing twelve months from the ones that do. Six tools:

Tool	Returns
trailing_financials	TTM revenue, EBIT, D&A and capex from the latest 10-K plus any newer 10-Q, with both accessions and the actual capital intensity ratios
market_quote	price, market cap, shares outstanding, enterprise value
market_beta	raw, Blume adjusted, and bottom-up sector beta
equity_risk_premium	Damodaran's current implied ERP, live, with its date and measure
risk_free_rate	US Treasury yield, 10y default
peer_metrics	peer multiples and margins, medians over complete rows only

No API key. yfinance and public data underneath.

Tools

tools/verify.py spawns both MCP servers exactly as dsh does, straight from the generated patch files, completes the handshake, lists the tools and makes one live call each. Run it after installing. A failed MCP mount is otherwise invisible: the child dies on startup, the tools are simply absent, and the model answers from memory instead of from filings with no error anywhere.

tools/recalc.py opens a workbook in Excel or LibreOffice so its formulas gain cached values, then verifies coverage and reports any #REF!, #DIV/0! or #VALUE!. openpyxl writes formulas without results, so without this step a generated workbook reads as empty to every validator downstream, including the one shipped inside dcf-model.

Why the market MCP does what it does

Two of its tools exist because the skills get these wrong, and the errors are large.

Beta. dcf-model says to use a five year monthly regression beta. Run that on NVIDIA in September 2026 and Yahoo returns 2.217, which produces a 17% WACC and a valuation 49% below the market price. That beta is measured across the period the stock rose roughly tenfold, so it captures the re-rating rather than systematic risk. Damodaran's semiconductor sector beta across 66 firms is 1.49. Relevered to NVIDIA's own capital structure it is essentially unchanged, because the company carries almost no debt, and the valuation lands within 10% of the market price. market_beta returns all three figures and recommends the bottom-up one, warning when the raw beta diverges from its sector by more than 30%.

Equity risk premium. The skill says "5.0-6.0% (market standard)" with no source and no measure named. Damodaran currently publishes five ERP estimates spanning 3.56% to 6.05%. His current implied figure is 4.14%; his ten year average cash flow yield is 6.05%. Both are defensible, they are different estimators, and on a high beta name the gap is worth tens of dollars per share. equity_risk_premium fetches the current implied figure live with its date and measure so the choice is recorded rather than assumed.

Base period. Nothing in the skills forces the trailing twelve month base, and dcf-model explicitly permits "LTM or most recent fiscal year". In a real run the agent had the newer 10-Q in front of it, mentioned the quarter's revenue 120 times, and still anchored on the fiscal year, which was two quarters stale and cost about a third of the valuation. trailing_financials assembles the base in one call so the correct choice is the easy one.

The check neither skill contains. Before reporting a valuation, solve for the discount rate the current share price implies. Where that differs from your WACC by more than about 300 basis points, the inputs are the more likely problem. That one step catches most bad cost of capital assumptions immediately.

Worked example

examples/nvda/ has a complete NVIDIA valuation: Excel model with 150 live formulas and zero errors, DOCX initiation report with 11 embedded charts, and the build scripts.

Requirements

Python 3.11+. Excel or LibreOffice for recalc.py. See requirements.txt.

Licence

Original work under MIT, see LICENSE. Vendored skills under Apache 2.0, see NOTICE and LICENSE-APACHE-2.0-anthropic.

Anthropic's xlsx, docx, pdf and pptx skills are not included. Their licence prohibits redistribution.

Not investment advice. The worked example demonstrates tooling and is not a recommendation on any security.