finances

See where your money goes from your own bank statements and exports. No bank login, no Plaid, no account linking, no cloud. Everything runs on your machine and nothing is uploaded anywhere.

uvx finances demo && uvx finances web --db demo.db --open    # try it on made-up data

or install it: uv tool install finances / pipx install finances / pip install finances.

20 seconds of 0.4, all on made-up data: the finances demo store and the reconciliation test fixtures. Click for the full video with sound.

The thing that makes it different: a statement prints its own totals, so the extraction can be proved correct. Every figure this produces traces back to a line on a statement, and the tool tells you when it cannot.

$ finances check

Store       finances.db
Rows        1180 transactions, 2025-01-03 to 2025-12-29
Statements  36

Reconciled  36/36 (100%)

Every statement reconciles and no months are missing. The totals downstream
are traceable to a statement line.

A tool built on a bank API cannot do this. It receives a feed of transactions with no independent total to check them against, so a dropped charge or a flipped sign is invisible. Here the statement is the ground truth, and reconciliation is enforced rather than assumed.

---

Running it

Install from PyPI, or use Docker. Either way you end up at the same dashboard on http://localhost:8787.

Try it first, with made-up data

Nobody should have to hand over a year of bank statements to find out whether a tool is any good. demo writes twelve months of an invented checking account and credit card — paychecks, rent, a subscription that raises its price, a holiday spike — into a separate demo.db. It goes through the same reconciliation and categorizer as a real statement.

# nothing to install permanently: uvx runs it from PyPI
uvx finances demo
uvx finances web --db demo.db --open

# or with Docker
docker compose --profile demo up demo        # then open http://localhost:8787

With your statements

Install it once. It needs Python 3.12+ (uv fetches it for you if needed); the dashboard comes prebuilt inside the package, so there is nothing else to set up.

uv tool install finances                 # or: pipx install finances
                                         # or: pip install finances

mkdir -p PDFs                            # drop statement PDFs and bank exports here
finances pipeline PDFs/                  # parse, categorize, check, report
finances import PDFs/                    # any CSV / OFX / CAMT.053 / MT940 files too
finances web --open                      # the dashboard

Upgrade with uv tool upgrade finances (or pipx upgrade finances). Your data is never inside the package: the store is finances.db in the folder you run from (or FINANCES_DB), and your statements stay wherever you put them.

Any folder layout works — subfolders are searched too. Tested against Amex, Chase (card and checking), Capital One and Discover. For any other bank, use its export (below), or let a local model read the PDF.

Always look at the check verdict at the end before trusting anything else. If it does not say 100%, the numbers downstream are built on a bad extraction, and every other report will tell you so too.

Any bank: CSV, OFX/QFX, CAMT.053, MT940

Nearly every bank lets you download your transactions. import reads those files, works out the columns, the date format, the decimal separator and which way the amounts are signed, and checks them against whatever balances the file carries:

finances import ~/Downloads/activity.csv
finances import exports/                 # a whole folder, any mix of formats

Format	Who exports it	How it is checked
CSV with a running balance	most bank accounts	every row has to move the balance by exactly its amount
CSV without one	most card sites	imported, and marked unverified — honestly
OFX / QFX / QBO	US banks ("Quicken" / "Money" download)	each export's opening balance must equal the previous export's closing balance
CAMT.053	European banks (ISO 20022 XML)	opening balance + every booked entry = closing balance
MT940	European banks (SWIFT)	opening balance + every entry = closing balance

Headers are recognised in English, German, French, Spanish, Italian, Dutch and the Nordic languages, 1.234,56 and 05.01.2026 are read the European way, and a summary block above the header (Bank of America does this) is skipped. When downloads overlap, rows already in the store for that account are skipped, not counted twice. If a CSV is ambiguous, say so: --sign out-negative, --date-format %d.%m.%Y, --account 1234 --issuer "My Bank" --kind card.

A bank the parser does not know: a local model, checked

finances parse PDFs/ --llm

When a statement's layout is not understood (nothing extracted, or the extraction does not reconcile), its text goes to a model running on your machine — Ollama, LM Studio or llama.cpp — which lists the transaction rows. The statement's printed totals are still read by the deterministic parser, so the model never supplies both the answer and the check: its rows are kept only if they add up to those totals. Otherwise they are discarded and the statement stays marked as a mismatch. FINANCES_LLM_FALLBACK=1 turns this on for the dashboard's uploads too.

With Docker

mkdir -p PDFs                            # drop your statement PDFs in here
docker compose up                        # then open http://localhost:8787

On start the container parses whatever is in ./PDFs (unchanged files are skipped), categorizes it, prints the check verdict and serves the dashboard. The store and every generated report land in ./data. Every command works the same way inside it:

docker compose run --rm finances check
docker compose run --rm finances cashflow
docker compose run --rm finances ask --example top-merchants

The port is published on 127.0.0.1 only, and .dockerignore is a whitelist: your PDFs and database are never sent to the Docker daemon or baked into an image.

The dashboard

Your statements, the reconciliation verdict for each one, spending by month and category, detected subscriptions, and a searchable transaction table. You can drag new statement PDFs or bank exports straight onto it: they are read on drop, and the reply tells you how many transactions came out and whether the file checks out.

• Wrapped — your year in money: where it went, your regular, the month that got away from you, what subscriptions cost you all year, and how much of it traces back to a statement that reconciles.
• Budgets & goals — a monthly limit per category, projected to the end of the month so you hear about it before it is broken; savings goals against the surplus your statements actually show.
• Cancel these — recurring charges ranked by what dropping them saves, with the reason each one is there (a price that went up, three services doing one job). Mark one cancelled and any charge after that date is flagged.

The server makes no outbound requests, binds to localhost, and refuses any request that does not come from a page on localhost — so another website open in the same browser cannot read your transactions from it.

Working on the code itself? See Developing below.

Or stay in the terminal

finances analyze            # where the money went
finances wrapped            # your year in money
finances budget             # this month against your limits
finances budget cancel      # the subscriptions worth cancelling
finances cashflow           # the next 60 days
finances audit              # the reckoning
finances cpi                # your own inflation rate
finances backtest           # how accurate are the forecasts
finances ask "how much did I spend on coffee in March"

finances on its own lists every command; finances <command> --help shows that tool's own options. If you did not install it, put uvx in front of any command: uvx finances check.

Configuration

Everything is a flag, and the few things you would otherwise repeat on every command are also environment variables:

Variable	Default	What it sets
FINANCES_DB	finances.db	the store every command reads and writes
FINANCES_PDF_DIR	./PDFs	where dashboard uploads are saved
FINANCES_HOST / FINANCES_PORT	127.0.0.1 / 8787	where finances web listens
FINANCES_CURRENCY	USD (or finances fx --base)	the currency every total is reported in
FINANCES_LLM_ENDPOINT	Ollama, LM Studio, llama.cpp on localhost	the model server for ask, enrich and parse --llm
FINANCES_LLM_FALLBACK	off	1 lets a local model read unknown PDF layouts, uploads included
FINANCES_PLAN	plan.json	where savings goals are kept
FINANCES_WEB_DIST	web/dist, or the copy inside the package	a prebuilt dashboard to serve

---

Command reference

Command	What it does	Writes
parse PDFs/	extract transactions from statement PDFs (--llm for unknown layouts)	finances.db
import exports/	bank CSV, OFX/QFX, CAMT.053 and MT940 files	finances.db
check	reconciliation and coverage report	—
categorize apply	label transactions, write back to the store	finances.db
analyze	spending dashboard	finance_dashboard.html, insights.md
cashflow	daily balance projection with a low-point warning	cashflow.html
plan status	savings goals against your real surplus	plan.json
budget	monthly limits per category; budget cancel for the cancel list	finances.db
wrapped	your year in money	—
export beancount	beancount, hledger/ledger, YNAB or Actual files	the file you name
mcp	let Claude and other AI assistants read the store (read-only)	—
verify	check statements and exports on their own, with an evidence report	the JSON you name
reconcile	match bank, processor and ledger records; explain what does not match	the JSON you name
evidence check	re-check an evidence report: digest, input files, re-run	—
fx	base currency and exchange rates	finances.db
audit	a strict grandfather goes through your statements	auditor_ledger.json
cpi	personal inflation, per merchant and per category	cpi.html
backtest	grades the forecasts against what actually happened	backtest.html
ask "..."	plain-English questions, answered offline	—
anonymize	a copy safe to share	transactions_shareable.csv
enrich	local-LLM merchant cleanup	finances.db
web	serve the browser dashboard, with PDF drag-and-drop	—
demo	a year of made-up statements, to try everything on	demo.db
pipeline PDFs/	parse + categorize + analyze in one run	all of the above

Every command accepts either the SQLite store or a CSV, interchangeably:

finances analyze --db finances.db
finances analyze cleanedData/transactions.csv

Things you will probably need

# Encrypted PDFs
finances parse PDFs/ --password 1234

# A statement extracted wrong — see what the parser saw
finances parse "PDFs/one bad file.pdf" --dump-columns

# Fail the build if anything does not reconcile (for a cron job)
finances check --strict

# Re-parse everything, ignoring the cache
finances parse PDFs/ --force

# Tell it your take-home pay, so it can work out what is left over
finances analyze --income 6200

# Teach the categorizer, biggest money first
finances categorize review

Hand it to the tool you already use

This gets complete, checked transactions out of statements. Budgeting apps and plain-text ledgers are better at other things, so it hands the rows over:

finances export beancount -o money.beancount --balances
finances export hledger   -o money.journal   --balances   # ledger-cli reads it too
finances export ynab      -o exports/        # one CSV per account
finances export actual    -o exports/

With --balances, every reconciled statement's closing balance becomes a balance assertion, so bean-check or hledger check re-verifies the whole history independently of this tool. Card payments from checking go through an Assets:Transfers account, so they net to zero instead of counting twice.

More than one currency

Each transaction keeps the currency its file was in. Totals are reported in one base currency, at rates you set — there is no rate feed, because nothing here goes online:

finances fx --base EUR            # report in euros
finances fx USD=0.92 GBP=1.17     # 1 USD is 0.92 of the base currency
finances fx                       # what is set, and what is still missing

Rows in a currency with no rate are left out of the totals and counted, and the dashboard says so, rather than quietly adding euros to dollars.

Verify and reconcile: the engine underneath

finances.verify is the part of this project other software can use. It reads bank statements and exports, Stripe payout reports and accounting ledgers, and checks whatever can be checked independently. It matches records across sources and writes an evidence report: every record says where it came from (page and line, CSV row, OFX FITID, camt reference), and every check and match says which rule decided it, on which inputs, with the expected and computed amounts.

finances reconcile --bank checking.csv --processor stripe_payouts.csv \
                   --ledger general_ledger.csv --ledger-account Checking \
                   --json evidence.json --sarif exceptions.sarif

Matching
  bank~ledger          7 matched, 1 needs_review, 3 unmatched
  bank~processor       2 matched

Needs a decision (1)
  bank      2026-10-06      252.00 USD  STRIPE TRANSFER
  ledger    2026-10-06      260.00 USD  Stripe payout
  > The difference, 8.00 USD, equals the processing fees on payout po_1Q8AbcD
    (3 transactions). The books show more than the bank received: the deposit
    looks recorded before fees.

• Matching runs in tiers: shared references first, then exact amounts within a date window (one-to-one, solved exactly), then grouped settlements (a deposit that paid several invoices, a payout made of many charges). Near misses come last. Those are never matched automatically: they go to review, with the difference explained when something known accounts for it.
• Each check gets one of five statuses (verified, mismatch, unverifiable, needs_review, incomplete), never a single score. A CSV with no balances is reported unverifiable, not verified.
• Exact money: Decimal and ISO 4217 minor units, with zero tolerance unless you set one.
• Reproducible: the same inputs and options give the same content digest. finances evidence check evidence.json --inputs ./ --rerun confirms the report is intact, that the files still hash the same, and that re-running reproduces it.

The digest proves the files have not changed since the run. It does not prove a statement is genuine.

From Python:

from finances.verify import VerificationEngine

eng = VerificationEngine()
eng.add("checking.csv", role="bank")
eng.add("stripe_payouts.csv")                    # recognised as a processor report
eng.add("general_ledger.csv", role="ledger", account="Checking")
report = eng.run()
for m in report.needs_review():
    print(m.pair, m.difference, m.explanation)
report.save("evidence.json")

The research behind the design (record-linkage theory, subset-sum limits, W3C PROV, in-toto, SARIF, audit evidence standards) and the roadmap are in docs/verification-engine.md.

Ask questions in the dashboard, with your own key

The dashboard's Ask page answers questions about your money with your own Claude or OpenAI API key, or with a model running on your own computer. Pick a provider, paste a key, and ask "where did my money go last month?" or "which subscriptions should I cancel?". Under each answer you can see which lookups the model made.

The model gets the same read-only tools as finances mcp below, so every number it gives you comes from your reconciled store, and it cannot change anything. The key stays in your browser: it is forgotten when you close the tab, unless you tick "Remember keys on this device". Each question carries the key to the local server, which passes it on to the provider and never writes it down. To use a key from the server instead, set ANTHROPIC_API_KEY or OPENAI_API_KEY before finances web. Setting OPENAI_BASE_URL points the OpenAI option at any server that speaks Chat Completions, such as Ollama, LM Studio, vLLM or OpenRouter.

The model menu lists each provider's current models, taken from their own model pages (Claude Opus 5.5, Fable 5.1, Sonnet 5.5 and Haiku 5.5; GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna). "Other model ID…" takes any model the menu does not list yet.

The conversation is kept with your data, so it is still there when you come back; the trash button under it deletes it, after asking once. Only the most recent turns are sent with each new question.

Local model: pick "Local model" and the page finds Ollama, LM Studio, a llama.cpp server or vLLM on this machine (or type its address), and lists the models it serves. No key is needed and nothing leaves your network; addresses outside it are refused. A model that cannot call tools still answers: the main figures are looked up first and handed to it with the question. Local models are slower; a 10B model takes about 15–20 seconds on a recent laptop.

With Claude or OpenAI, your question and what the tools return (totals, merchant names, amounts) are sent to the provider so it can write the answer. Your statements and database stay on your machine.

Ask Claude (or any AI assistant) about your money

finances mcp lets an AI assistant read your store through MCP, so you can ask in plain English, inside an app you already use:

• "Where did my money go in March compared to February?"
• "Which subscriptions went up in price this year, and what would cancelling them save?"
• "Can I afford a $2,000 trip in January without dropping below $1,000 in checking?"
• "Draft an email disputing the duplicate $89 charge from Oct 14."

The assistant calls tools like spending_by_category, search_transactions, subscriptions, cancel_list, cashflow_forecast, anomalies, wrapped and check, and answers from the reconciled numbers instead of guessing. There is also a read-only query tool that takes a SQL SELECT, with the same guard finances ask uses. Nothing it does can change your data, and getting data in works exactly as before: statements and exports you download yourself.

Claude Desktop: add this to claude_desktop_config.json (Settings → Developer → Edit Config), using the absolute path to your store, then restart Claude:

{
  "mcpServers": {
    "finances": {
      "command": "uvx",
      "args": ["finances", "mcp", "--db", "/Users/you/finances/finances.db"]
    }
  }
}

Claude Code:

claude mcp add finances -- uvx finances mcp --db /Users/you/finances/finances.db

Any other MCP client (Cursor, Windsurf, LM Studio, …) takes the same command.

Know where the answers go. Your store stays on your disk, but a cloud assistant (Claude, ChatGPT, …) receives what each tool returns (totals, merchant names, amounts) in order to write its reply. To keep everything on your machine, use an MCP client running a local model, such as LM Studio.

Asking questions offline

ask needs a local model server. Without one, --sql and --examples still work:

finances ask --examples
finances ask --sql "SELECT category, SUM(amount) FROM transactions GROUP BY 1 ORDER BY 2 DESC"

With one:

brew install ollama && ollama pull llama3.1:8b
finances ask "what did I spend on delivery last month" --explain

Something in the 3B–8B range is plenty. Avoid reasoning models here — they take minutes per question for no benefit on this task. Pick explicitly with --model llama3.1:8b.

If something goes wrong

Symptom	Cause and fix
finances.db does not exist	Run parse first.
check reports mismatches	The extraction is wrong for those files. Run finances parse <file> --dump-columns to see the columns it found, or --llm to let a local model try.
A bank is not supported	Download its CSV/OFX export and finances import it, or ask for the bank.
check reports gaps	A statement month is missing from your folder. Totals are being computed on incomplete data.
Lots of Uncategorized	categorize review to teach it, or enrich to use a local LLM.
cashflow will not project	It could not find a checking-account balance. Pass --balance 2400.
Amounts look wrong	Almost always a column-detection problem. --dump-columns, then --amount-col last or second-last to override.

---

How it fits together

PDFs ─────▶ parse.py ───┐
   (llm_parse.py, checked)  ├──▶ finances.db ──▶ engine.py ──▶ every other tool
exports ─▶ importer.py ─┘          │                 │
                                   │          the read model: what counts as
       each file checked against   │          spending, which months are complete,
       its own totals or balances  │          what repeats, what is verified
                                   └──▶ export.py ──▶ beancount, hledger, YNAB, Actual

finances/          the package — one module per tool, plus the shared core below
  cli.py           `finances <command>`: the front door, dispatches to each tool
  parse.py         PDFs -> the store, reconciling each statement as it goes
  importer.py      bank exports (CSV, OFX, CAMT.053, MT940) -> the store, balance-checked
  llm_parse.py     a local model for unknown PDF layouts, kept only if it reconciles
  engine.py        the read model every tool shares
  export.py        the store -> beancount, hledger, YNAB, Actual
  mcp_server.py    the store as read-only tools for AI assistants (MCP over stdio)
  chat.py          the dashboard's Ask page: your own Claude or OpenAI key, read-only tools
  verify/          the verification engine: adapters, checks, matching, evidence reports
  api.py           the dashboard's JSON API and static server
  demo.py          the made-up store
  ...              analyze, cashflow, cpi, backtest, ask, planner, anonymize, ...
web/               the dashboard (React + Vite), served by `finances web`
tests/             stdlib unittest, no test dependencies
docker/  Dockerfile  docker-compose.yml

Every module still runs on its own too: python -m finances.parse --help.

• models.py — one definition of a transaction and a statement. The SQLite schema and the CSV columns are both generated from it, so adding a field is a single edit and the store migrates itself.
• database.py — the store, plus a content-hash cache so re-running the parser skips files that have not changed.
• utils.py — money, date and merchant parsing, the debt formulas, and the recurring-charge detector.
• engine.py — the semantics. Everything downstream agrees about what a month, a surplus and a subscription are because they all ask this.

What it does that other tools do not

Proves its own numbers. Extracted charges are checked against the total the statement prints for itself, and the verdict is stored, not just displayed. Missing statement months are detected and reported as gaps, so an average is never quietly computed over incomplete data.

This is not decorative. Building it found a noise filter matching the month "Apr" and discarding entire statements, sign inversions on two issuers, a year-rollover that dated 62 transactions to the wrong December, and a primary key that silently dropped 41 of one statement's 181 transactions. All of them looked fine in a dashboard. None survive a reconciliation check.

Tells you about next month, not just last month. cashflow projects the daily balance forward from known recurring charges and detected paycheck cadence — distinguishing semi-monthly from biweekly, because biweekly gives you three-paycheck months and that changes the answer. It reports a band rather than a line and says plainly that it is not a confidence interval.

Measures your inflation, not the country's. cpi weights per-category price changes by your spending and tracks per-merchant prices over time, against real BLS Consumer Expenditure and CPI-U reference data shipped with the project.

Grades its own forecasts. backtest walks forward month by month, predicts each month using only data from before it, and scores nine predictors against what happened. It refuses to name a winner when the sample cannot support one.

Makes your data shareable. anonymize has four levels up to fully synthetic, and proves what survived — per-category totals are printed side by side so you can see which figures are still exactly true.

Categorization

Rules, in order, first hit wins. Every answer records which rule fired and how confident it is:

1. what you taught it (learned_rules.json, always wins)
2. context rules — COSTCO+GAS is fuel, UBER+EATS is delivery
3. bank phrases — a deposit account names its own transaction type
4. brand table, exact token match, never substring
5. keyword rules, word-boundary regex
6. issuer MCC labels, where the statement prints its own category
7. behaviour — cadence and amount, when the name says nothing

Scored against 159 real-shaped descriptors with processor prefixes, glued city/state suffixes and issuer preambles:

$ finances categorize eval
  correct         159  100.0%
  wrong             0    0.0%   <- these corrupt analysis
  left unknown      0    0.0%   <- these are honest

Developing

git clone https://github.com/roshanlam/finances && cd finances
uv sync                                  # editable install into .venv
uv run finances demo                     # `uv run` uses the checkout's code
uv run python -m unittest discover tests

In a checkout the dashboard needs Node 20+: uv run finances web builds it on first run. While working on the frontend, run uv run finances web in one terminal and cd web && npm run dev in another, then use http://localhost:5173 (it proxies /api through).

Releasing a new version:

# bump `version` in pyproject.toml and add a CHANGELOG.md entry, then
rm -rf dist finances/web_dist
(cd web && npm ci && npm run build) && cp -r web/dist finances/web_dist
uv build && uv publish

344 tests, no test dependencies. The parser's golden-file tests skip when PDFs/ is empty, so a fresh clone runs the rest. The golden-file harness records per-statement extraction totals so a parser change that moves a number is visible:

uv run python tests/run.py --update-golden   # after a deliberate change
uv run python tests/run.py --check-golden    # exits 1 on drift

Privacy

Your statements and the store never leave the machine. PDFs/, *.db, the generated dashboards and anonymize_map.json are all gitignored. The map file is what turns MERCHANT_007 back into a real name — keep it, do not share it.

ask, enrich and parse --llm talk only to a model server on localhost. Nothing else in the project makes a network request at all. finances mcp makes none either, but whatever assistant you connect to it sees the results of the tools it calls; see Ask Claude. See SECURITY.md for how the dashboard is locked to your own machine and how to report a problem.

Not financial advice

Arithmetic on your own statements. The forecasts are extrapolations of your past spending and say so, backtest will tell you how wrong they have been, and none of it knows anything about your life beyond what the statements print.

License

MIT. See LICENSE.