finkrit 0.3.1

Portfolio risk, performance, optimization, and tax analytics. An open core quant engine, with an optional conversational agent layer and a web dashboard on top.

Ask a real portfolio question

Which holdings in my portfolio are the riskiest?

finkrit separates the dimensions of risk instead of forcing them into one score. In the bundled portfolio, NVDA has the highest annualized volatility at 38.05%, while MSFT has the deepest maximum drawdown at -34.50%. AMZN is next by volatility at 33.94%. The agent explains the distinction and shows the metrics behind the answer, so you can decide whether volatility, downside history, or both matter for the decision at hand.

https://github.com/user-attachments/assets/aa26a18a-dc8f-4f3e-8dc2-f1833fb9baaf

Try it on the bundled example

No file of your own yet? example loads a sample that ships with finkrit, so this works straight after install:

finkrit cli --file example

# against a local model, no key needed
finkrit cli --file example --url http://localhost:11434/v1 --model qwen2.5:14b-instruct

# from a source checkout, where finkrit is not on the path
./run cli --file example --url http://localhost:11434/v1 --model qwen2.5:14b-instruct

It opens by asking which agent to use, then prints what it parsed so you can check the file was read the way you meant. One row per tax lot, not per holding, which is the distinction the whole design turns on:

Then ask it things. Pick 0 for the router and it works out which specialist each question belongs to, showing the sub question it delegated, the tool that ran, and the answer that came back:

Note what the answers state without being asked: that beta is unitless, that CVaR is a fraction of value rather than a currency amount, the exact window the figures cover, and which benchmark was used. Those come from the tool results rather than from the model, which is what stops a small model reporting a fraction as a dollar figure.

It is a twelve position portfolio built from sixteen tax lots, formatted the way a custodian actually exports: dollar signs, quoted thousands separators, and MM/DD/YYYY dates. Three names were bought more than once, which is the part worth paying attention to.

Ask it this:

Which of my lots are sitting at a loss, and what could I harvest?

AAPL is the case the whole lot level design exists for. It is one holding of 180 shares, and as a single blended position its cost basis is $27,431.50, about $152.40 a share. At that average the position looks like a straightforward winner and there is nothing to harvest. But it is really three purchases:

Lot	Quantity	Cost / share	Acquired
1	100	$120.40	2021-05-12
2	50	$180.15	2023-03-09
3	30	$212.80	2024-06-03

The 2024 lot cost nearly twice what the 2021 lot did. Whenever AAPL trades between those two numbers, that third lot is underwater while the position as a whole is up, and it is harvestable even though the holding is profitable. Averaging the lots together makes that loss invisible. UNH has the same shape at $412.60 against $492.30, and MSFT at $238.60 against $362.45.

Those cost figures come from the file and never change. What the lots are worth today depends on live prices, so the answer moves with the market.

Worth also trying:

How much of this portfolio qualifies for long term treatment?

What is my volatility, and which holding contributes most to it?

The second one fans out to more than one specialist. In the dashboard you can click each specialist's name on the reply to see exactly what it returned before the answers were combined.

A CSV file has one row per tax lot, with four columns: ticker, quantity, cost per share, and acquired date. For example:

ticker	quantity	cost_per_share	acquired
AAPL	100	120.00	2021-05-12
AAPL	50	180.00	2023-03-09
MSFT	95	238.60	2021-02-18
NVDA	140	168.20	2023-03-09

Repeat a ticker for each time you bought it. AAPL above is one holding of 150 shares made of two lots, and they stay separate all the way through. That matters for tax, because a position can be up overall while individual lots are underwater, and those are the ones worth harvesting. Blending them into one average cost hides exactly the losses you are looking for. Buy once and a single row is all you need.

Column names are matched case insensitively against common aliases, so a typical brokerage export loads without renaming anything:

Field	Recognized column names
Ticker	ticker, symbol
Quantity	quantity, shares, qty, units
Cost per share	cost_per_share, cost per share, cost/share, cost basis / share, cost basis per share, price per share, cost basis, avg cost, average cost basis, cost, price, price paid
Acquired	acquired, date acquired, purchase date, date
Name (optional)	description, name, security, security name, company, company name

The first four are what a file must label. Name is read when it is there and skipped when it is not, so a file without it still loads. It is worth having: most exports print the security next to its symbol, and an agent handed a bare ticker will supply a company name from memory and can get it wrong.

Dates accept YYYY-MM-DD, MM/DD/YYYY, MM/DD/YY, or DD-MM-YYYY. Commas in numbers are stripped, extra columns are ignored, and a missing or unreadable date falls back to a default.

The web upload uses that same table. When your header names all four fields under any of the spellings above, the file is read in code: instantly, with no model involved and no key needed. Only a file that leaves one of the four unnamed goes to the model, which maps whatever columns and formats it finds onto the same four fields and flags anything it had to guess. So almost any layout works, and a tidy one costs nothing.

What differs between the two is the response to a gap. The terminal substitutes a default and carries on, since a chat session is throwaway. The upload records it on the holding for you to correct before anything is saved.

The web app

finkrit with no subcommand builds the dashboard, serves it, and opens your browser at http://127.0.0.1:8000:

finkrit

# on a local model, no key
finkrit --url http://localhost:11434/v1 --model qwen2.5:14b-instruct

# from a source checkout
./run --url http://localhost:11434/v1 --model qwen2.5:14b-instruct

Your book, lot by lot

Upload a CSV and it parses into positions and the lots underneath them. The overview counts positions rather than rows, allocation is by cost basis, and every field in the table is editable in place before you commit it with Save portfolio.

Risk

Volatility, value at risk, beta and maximum drawdown, computed in code with no model anywhere in the request. The line under the cards states the as of date, the sampling interval and that these are log returns, so the numbers are never floating free of the window they came from. First load fetches prices for every holding plus the benchmark and reports progress per ticker.

Tax signals

Every lot trading below cost and clear of the wash sale window, priced at your assumed rates so the saving is a dollar figure rather than a hint. Each card names the lot, what it cost, what it is worth now, and what realizing it would save. As the footer on that screen puts it, the signals are computed from your lots in code and nothing there comes from a language model.

Rebalance

Three strategies run against the same target and the same budget, so the tradeoff between tax paid and drift left is a table instead of an argument. With no gain budget, every overweight sells the whole way and the tax bill is whatever it is:

Set a budget and the plan spends it to the dollar. The same comparison at $5,000 sells two names instead of five, names the three it deferred, and the residual drift jumps from 1.96% to 21.69%, which is the price of the smaller tax bill stated rather than implied:

Ask it anything

The chat panel opens beside whatever you are looking at and routes through the orchestrator, so a question reaches the right specialist without you picking one. The pill above each reply names the specialist that answered, and clicking it shows what that specialist returned before the answers were combined.

Asked in a fresh turn, the CVaR answer states the figure, that it is a fraction of value rather than a currency amount, the confidence level, and the exact window it was computed over. None of that came from the model. It came from the tool result:

What is in here

finkrit is a small, layered stack, a quant core with Agentic AI, an API, and a web app built on top.

Path	Import name	What it does
packages/finkritq	finkritq	Deterministic quant core. Holdings, tax lots, prices, risk, performance, optimization, and tax. No agent or web dependency.
packages/finkritintel	finkritintel	Tool contracts and capabilities. The bridge that exposes the core as callable tools, framework neutral.
packages/finkritcore	finkritcore	Deterministic domain layer. The store, the report composers, and the CSV mapper, tied together behind a Desk. No agent framework, so the dashboard and the reports run without one.
packages/finagent	finagent	Conversational agents over the capabilities, built on pydantic-ai.
services/api/finkritserver	finkritserver	FastAPI layer that serves the JSON API and the built web app.
apps/finkritweb	(web)	SvelteKit dashboard. Upload a portfolio, see it, ask about it.

finkritq is the open core and stands on its own. Everything above it adds tools and an agent, and stays optional.

The layering is a graph rather than a stack. finkritintel and finkritcore are siblings on top of finkritq, answering two different questions: intel is what the system can compute, core is what a client holds. finagent stands on both. tests/test_layering.py walks the shipped source and fails the suite if a lower layer ever imports an upper one, so the rule is enforced rather than described.

Quickstart

pip install finkrit           # or: pipx install finkrit
export LLM_API_KEY=sk-...      # any OpenAI, Anthropic, or Google key
finkrit                        # start the dashboard, opens your browser

finkrit wants an LLM key for the chat. The dashboard, the risk report, and a CSV upload whose header names its columns all run without one. Any provider pydantic-ai supports works, keyed by the single LLM_API_KEY variable, or pass it inline with finkrit --key sk-....

Prefer the terminal? finkrit cli chats with the agent over a portfolio instead.

Command line

finkrit            start the dashboard (opens your browser)
finkrit web        the same, explicit
finkrit cli        chat with the agent in the terminal

The dashboard takes:

finkrit --key sk-...           the LLM key inline (should match provider)
finkrit --model openai:gpt-5   pick the provider and model (defaults to openai:gpt-5)
finkrit --url http://host/v1   run against a local model, no key (see Local model)
finkrit --port 8001            serve on a different port
finkrit --dev                  Vite hot reload (source checkout only in case you want to tinker)

Local model

Point the agent at any OpenAI compatible endpoint (a local Ollama, LM Studio, vLLM, llama.cpp server, or a self hosted box) with --url. No cloud key is needed, and you set the model to whatever the endpoint serves:

# a local Ollama, the usual case
finkrit --model qwen2.5:14b-instruct --url http://localhost:11434/v1
finkrit cli --model qwen2.5:14b-instruct --url http://localhost:11434/v1

# a box elsewhere on your network. YOUR-HOST is a placeholder, substitute it
finkrit cli --model qwen2.5:14b-instruct --url http://YOUR-HOST:8000/v1

From a source checkout, where finkrit is not on the path, ./run takes the same arguments:

./run cli --model qwen2.5:14b-instruct --url http://localhost:11434/v1 --file example

--model names the flag on both entry points. Behind --url it is the name the endpoint serves, verbatim, so an Ollama tag like qwen2.5:14b-instruct goes through whole rather than being read as a provider prefix.

The agent leans on tool calling, so use a tool capable model (llama 3.1 or 3.3 70B, qwen2.5-instruct, and similar). Small models often fumble the tool calls. And nothing leaves your machine, so a local model keeps the whole conversation private.

Chat with the agent

finkrit cli is a REPL over a portfolio. With no --file it uses a seeded offline portfolio, 40 AAPL, 30 MSFT, 20 NVDA, 25 JPM, and 35 XOM, each at a cost basis of 100 acquired 2022-01-03, priced with deterministic fake data so runs are reproducible. Point it at your own holdings with a CSV instead, which switches to live market data:

finkrit cli --file my_holdings.csv

Every flag the CLI takes:

-f, --file PATH      a portfolio CSV, or 'example' for the bundled sample.
                     Left off, the seeded offline portfolio is used
-ag 0|1|2|3|4        router, risk, optimization, performance, tax. --agent is
                     the long form, and the name works in place of the number.
                     Left off, a menu asks
--model NAME         provider shortcut (claude, openai, gemini, groq, mistral),
                     a full provider:name string, or the served name behind --url
--url URL            an OpenAI compatible endpoint, a local Ollama, LM Studio,
                     vLLM or self hosted box. No key needed
--key sk-...         the LLM key inline, instead of the LLM_API_KEY env var
--lang NAME          language to answer in, English by default
--steps              also show each tool's arguments and each specialist's answer
--truncate-steps     cut each step to one terminal row, for a narrow terminal
--quiet              hide the live step trace entirely
--logs               print finkritq's data fetch logs, every download and cache hit

The agents

Under the chat sit five agents, four specialists and a router. Each specialist owns one domain, and no tools outside it.

Agent	Answers	Covers
Risk	how risky, what could be lost	volatility, variance, semivariance, downside deviation, drawdown and maximum drawdown, value at risk and conditional VaR, beta, and each holding's marginal and component contribution to risk
Performance	how it has done	total return, annualized return, and the risk adjusted Sharpe, Sortino, and Calmar ratios
Optimization	what to hold	the minimum variance and maximum Sharpe target weights, long only, and a tax aware rebalance plan toward them: sells chosen drift first, lots picked to minimize the gain, capped by a capital gains budget. Can compare strategies side by side, selling fully to target, just to the band edge, or partially filling to exactly spend the budget, each with its tax cost and remaining drift. Proposed allocations and plans, never trades
Tax	what the IRS sees	unrealized gains and losses per lot, tax loss harvesting candidates net of the wash sale window, and the long versus short term split. Read only, describes the tax position and never trades
Orchestrator	anything, mixed	reads the question, calls whichever specialists can answer, and combines their replies into one

Target a single specialist with -ag, by number or by name:

finkrit cli -ag 1                # risk
finkrit cli -ag performance      # same as -ag 3
finkrit cli --agent optimization

Numbers are 0 router, 1 risk, 2 optimization, 3 performance, 4 tax. Left off, the CLI shows a menu. A single specialist is the direct path, the model sees that domain's tools and answers with no routing overhead, so pick one when you already know the domain.

How the orchestrator works. The router (agent 0) is itself an agent whose tools each hand a focused sub question to one specialist. It reads your question, decides which specialists it needs, calls them (one or several), and synthesizes a single answer. So a mixed question in one message, for example "what is my volatility, my annualized return, and the optimal weights?", fans out to those three and comes back combined. It never invents or alters a number, and it reports what a specialist returned and nothing besides. The tradeoff is one extra model loop around the specialists it invokes, which is why a single specialist is cheaper when the domain is known.

When exactly one specialist answered, its reply reaches you word for word rather than being rewritten. There is nothing to combine in that case, so the orchestrator's closing text would be a second pass over numbers it was told not to touch, which is where two observed failures came from: a beta restated as -0.05 where the specialist had said -0.06, and a Chinese answer summarized into Thai while the model was instructed in English. A question that genuinely spans specialists is still combined, because there the synthesis is the work.

The web dashboard always routes through the orchestrator, so any question, risk, performance, or allocation, reaches the right specialist without you choosing one.

Follow up questions

Both the terminal and the dashboard carry the conversation forward, so pronouns and refinements work:

> betas of my holdings
> and in dollars?
> which of those is riskiest

History is the message thread of the top level agent. A specialist's internal tool loop happens inside a single tool call and never enters it, so the thread reads like the conversation you actually had. It is capped at forty user turns and trimmed at turn boundaries, never between a tool call and its result.

Where the numbers come from

Two rules hold the whole stack up, and both are enforced rather than requested.

The model may route, read and explain, but it never authors a number. Every figure comes back from a finkritq function it can call and cannot write.

Then the answer is read back before you see it. Each figure in the reply has to be a faithful rendering of some number the run actually received, where faithful allows rounding, converting a fraction to a percentage, and thousands separators, and does not allow a digit that changes. An unsupported figure is sent back to the model with the offending numbers named. If it will not correct itself the answer arrives annotated rather than silently wrong.

From source

To hack on finkrit, clone it and use the bootstrap, which sets up a virtual environment, installs dependencies, builds the web app, and launches.

git clone https://github.com/finkrit/finkrit
cd finkrit
export LLM_API_KEY=sk-...
./run                          # same flags as finkrit web, for example --dev

Prerequisites: Python 3.11 or newer and Node 18 or newer. Later runs skip the setup.

./run takes the same subcommands the installed finkrit command does, so a checkout reaches the terminal chat without installing anything:

./run                            # the dashboard
./run cli --file example         # the terminal chat, against your working tree
./run cli --file example --url http://localhost:11434/v1 --model qwen2.5:14b-instruct

Every flag documented above works after the subcommand, and ./run always runs the code you are editing rather than the published wheel.

Note that finkrit itself is not on the path in a fresh checkout, since the bootstrap installs dependencies rather than the project, and activating the environment does not change that. The error is a bare command not found. Use ./run as above, or put the command on the path with an editable install:

source .finkritvenv/bin/activate
pip install -e .
finkrit cli --file example

One thing the editable install does not cover: finkritintel and finkritq still come from PyPI, so finagent and finkrit are what track your edits. Change either of the lower two packages and you are testing the published version of it, not yours.

Using the quant core on its own

finkritq is the open core, published on its own so you can install just the quant engine without the agent or web layers.

pip install finkritq            # core, numpy and scipy only
pip install "finkritq[data]"    # adds the live yfinance data provider

It also ships a runnable demo that prints every analytic pillar over a portfolio, no agent involved:

python -m finkritq                                             # seeded, offline
python -m finkritq real NVDA KO PG --benchmark SPY --years 3   # needs [data]

Development

Sources live under packages/ and services/api/. The test runner is configured to put those on the import path, so a fresh clone runs the suite with no extra setup.

pip install -r requirements.txt
pytest                       # the whole suite
pytest packages/finkritq     # one package

Status

Early and moving. The layers above finkritq are the newest. Expect the agent and web surfaces to change while the core settles.

Disclaimer

finkrit is for educational and informational purposes only. It is not financial, investment, or tax advice, and nothing it produces is a recommendation to buy or sell any security. Use your own judgment and consult a licensed professional before making decisions.

The optional data provider uses yfinance to pull market data from Yahoo Finance. finkrit and yfinance are not affiliated with, endorsed by, or sponsored by Yahoo. That data is subject to Yahoo's terms of use and is intended for personal and educational use. Verify anything you rely on against an authoritative source.

The software is provided as is, without warranty of any kind.

License

Apache-2.0. See LICENSE.