@evan-moon/firma

firma

You earned well this year.
Do you know where it went?

Drop your trade history into Claude.
Get back a full financial picture — portfolio, net worth, cash flow — stored in a local SQLite file.

npm install -g @evan-moon/firma

You:    Here's my IBKR trade history [trades.csv]. Set up firma.
Claude: I see 47 transactions. Logging in chronological order... done.
        Syncing prices and FX history... your portfolio: $179K, +61% all-time. TSLA is 78%.

You:    My portfolio feels really volatile lately. Should I be worried?
Claude: TSLA is 74% of your equity — a 10% TSLA move is a 7% net worth move.
        Cash flow is solid at $6.4k/month saved. The volatility is market, not lifestyle.

You:    That July expense spike looks alarming.
Claude: That was $57k in TSLL — a leveraged bet, not spending.
        Strip it out: expenses flat at ~$7k/month, savings rate 47.8%.

You:    Give me a full overview with charts.
Claude: Net worth: $294k → $490k in 8 months, pulled back to $427k with TSLA's Q1 dip.
        [builds net worth trend, income vs expenses, portfolio allocation, unrealized P&L]

Why firma

• Talks to Claude natively. firma mcp install wires firma into Claude Desktop. Every command you can run in the terminal, Claude can call through MCP — log trades, sync prices, render charts, reason about your data.
• Local-first, no exceptions. Everything lives in ~/.firma/firma.db. No server, no sync, no account. Your financial data never leaves your machine.
• Multi-market, one stored currency. Track holdings across 9 markets (US, KRX, KOSDAQ, JP, HK, LSE, XETRA, NSE, TW) — plus non-stock assets like crypto, commodities, and real estate. Every value is stored in USD and displayed in the currency you pick, with historical FX rates from FRED — so a 2018 balance shows the 2018 exchange rate, not today's.
• Transactions as source of truth. No holdings table to drift out of sync. Portfolio, cost basis, and P&L are always derived live from your trade log.
• A daily brief that actually advises. get_market_brief doesn't just dump data — it returns a weekly stance (lean aggressive or defensive), a ranked recommendation per holding (BUY / ADD / HOLD / TRIM / SELL / AVOID with conviction, time horizon, English thesis, and a show_valuation drill-down), and a watchlist of new-ticker candidates, all computed deterministically from your holdings, cash flow, and macro signals. Each holding is auto-classified into one of five categories (growth, quality, hedge, income, speculative) and run through a category-specific playbook that enforces position-size ceilings. A cross-cutting correlation guard then downgrades any BUY/ADD that would lift the portfolio's weighted-average 90-day pairwise correlation by more than 0.10 and push it above 0.70, so "more of what you already own" gets caught before it lands. The same call bundles macro signals, commodity prices, the dividend calendar, global macro (World Bank), and active disaster alerts (GDACS) — no extra API keys needed for the global data.
• A simulator nobody else ships. project_net_worth runs a multi-scenario Monte Carlo (1000-iteration GBM, shared seeded shocks across scenarios) over your portfolio at any horizon — next month for short-term outlook or 30 years for FIRE planning, one tool. You supply the (return, vol) assumptions for each scenario and Claude discloses them verbatim; firma never bakes in defaults. Returns P10 / P25 / P50 / P75 / P90 paths, probability of reaching an optional target, and median months to reach. Diagnostic and thought-experiment tools stay strictly separated — the brief never invents future numbers, projections always show their assumptions. (See Forward projection below.)
• Time-series features work on day one. The first firma sync after install backfills 90 days of daily-close history from Yahoo's keyless chart endpoint — full shares × close snapshot rows for everything you hold and price-only rows for watchlist peers. Correlation, effective_n, Sharpe, max drawdown, annualized volatility, and CAGR are populated immediately, instead of waiting 30+ days for snapshots to accrue. Subsequent syncs skip tickers that already have coverage, and a single ticker failing on Yahoo never poisons the rest of the run.
• Fund-manager voice, not data dump. Claude talks back like a private fund manager — decisive ("MSFT ADD high-conviction, not before Thu PCE"), evidence-based (cites the firma fields behind every call), micro-macro bridged (every macro statement ends with a dollar number on your book, not "could pressure tech"). The brief's recommendations layer + a Voice-and-judgment instruction layer keeps the answer at the level a paid advisor would give — and refuses to fabricate when data is missing.
• Tracks your decisions, not just your positions. show_thesis_track joins your recorded reason per trade to current outcomes (validated / underwater / cut_loss). show_anti_portfolio follows the post-exit prices of positions you fully sold (missed_rebound / good_call) — the data brokerages deliberately don't show, because it highlights bad sells. Six months from now, Claude reads your trade reasons back to you with proof.
• Two clean tool families. Diagnostic (get_market_brief, every show_*, report) returns currently true facts — no future-looking probabilities. Thought-experiment (project_net_worth, simulate_scenario) returns conditional outcomes only when you supply assumptions, and Claude discloses them verbatim. The brief never invents future numbers; projections always show their math.

Get started

Three ways to install — all work the same way once set up.

macOS Installer

Download and double-click. Node.js is bundled — nothing to install first. Claude Desktop and Claude Code are connected automatically; a setup dialog then shows the two firma config set commands for your free API keys.

↓ Download firma-mac-arm64.pkg

Windows Installer

Download and run. Node.js is bundled and firma is added to your PATH — nothing to install first. Claude Desktop and Claude Code are connected automatically; a setup dialog then shows the two firma config set commands for your free API keys. The installer is unsigned — if SmartScreen warns, choose More info → Run anyway.

↓ Download firma-win-x64.exe

npm

# 1. Install
npm install -g @evan-moon/firma

# 2. Set API keys (both free)
firma config set finnhub-key YOUR_KEY   # finnhub.io — prices, news, earnings
firma config set fred-key YOUR_KEY      # fred.stlouisfed.org — macro & FX history

# 3. Connect Claude Desktop
firma mcp install
# Restart Claude Desktop — firma tools will appear in the toolbar.

Drop your data into Claude. Paste a CSV, brokerage export, or plain text — Claude logs everything and syncs prices automatically.

Prefer the terminal? The CLI is the safety net — setup, manual entry, repair, and a few show commands for quick inspection. Analysis, briefs, and projections go through Claude. See CLI reference.

If firma helps you track your money, please ⭐ star the repo — it's the cheapest way to help others find it.

CLI reference

The CLI is the safety net — three things only: setup, manual data entry, and repair. Anything analytical (brief, recommendations, projections, news, valuation, macro) goes through Claude.

Setup

Command	What it does
firma config set finnhub-key KEY	Set Finnhub API key (prices, news, earnings)
firma config set fred-key KEY	Set FRED API key (macro & FX history)
firma config set currency CODE	Set display currency (USD, KRW, JPY, …)
firma config get [key]	Show config values
firma mcp install	Register MCP server in Claude Desktop
firma doctor	Check API keys, data state, FX cache, time-series coverage
firma profile	Set up your goals (birth year, retirement target, target net worth, risk tolerance, notes) — every field optional

Manual data entry

Command	What it does
firma add txn	Record a transaction — a stock in any of 9 markets, or a non-stock asset (crypto / commodity / real estate) — type buy / sell / deposit / dividend / tax
firma add balance [-p YYYY-MM]	Monthly asset & liability snapshot
firma add flow [-p YYYY-MM]	Monthly income & expense entry
firma add monthly [-p YYYY-MM]	Balance + flow in one call (month-end)
firma add snapshot	Sync prices and record today's portfolio snapshot
firma add price [name] [value]	Set the current value of non-stock holdings (they have no price feed)

Repair (when Claude or you got something wrong)

Command	What it does
firma edit txn / balance / flow / snapshot	Edit an existing entry (interactive picker)
firma delete txn / balance / flow / snapshot	Delete an entry. Alias: firma rm

Verify (quick inspection)

Command	What it does
firma show portfolio	Holdings table with P&L, avg cost, market value, sector/country allocation, FX return decomposition (for non-USD display currency)
firma show txns [ticker]	Transaction history
firma show balance [period]	Stored balance entries
firma show flow [period]	Stored cash flow entries
firma show snapshot [ticker]	Portfolio value history; --from/--to for date range
firma show profile	Show the stored profile
firma show fx [currency]	Inspect cached FX history with --from/--to/--limit

Sync

Command	What it does
firma sync	Fetch latest prices (Finnhub) + FX rate history (FRED). On the first run, also pulls 90 days of Yahoo daily-close history for held stocks and watchlist peers so correlation / risk metrics work immediately. Idempotent — later runs skip tickers already covered
firma sync fx	FX history only — incremental backfill from your earliest entry date

Claude integration (MCP)

After firma mcp install, Claude has full read/write access to your data through conversation. Tools split into two epistemic families — Claude is told never to mix them:

• Diagnostic — get_market_brief, every show_*, report. Returns currently true facts: deterministic, rule-derived, no forward-looking probability distributions. The brief never invents future numbers.
• Thought experiment — project_net_worth (stochastic Monte Carlo distribution over months/years) and simulate_scenario (deterministic conditional point-estimate for shocks within days/weeks). Caller must supply assumptions (return + vol, or shock magnitudes), and Claude discloses them verbatim. Firma never bakes in defaults.

Voice and judgment. On top of the data and tools, firma ships a fund-manager voice layer in its server instructions: 4 explicit principles that govern how Claude answers — decisiveness over neutrality, every macro statement must end with a dollar number on your book, behavioral anchors (show_thesis_track on buys, show_anti_portfolio on sells), and every BUY/HOLD/SELL stance must cite the firma field that justifies it. Vague hedge phrases like "could pressure tech" or "조건부로 보라" are explicit violations. This is why firma's answers read like a paid advisor's, not a chatbot's data dump.

get_market_brief is the primary diagnostic entry point. In a single call it returns:

• Stance — a deterministic weekly verdict (aggressive / mixed / defensive / contrarian-aggressive) at the intersection of market signal and your personal capacity (runway, goal pace, idle cash)
• Recommendations — a per-ticker recommendations[] array, one entry per holding, with action (BUY / ADD / HOLD / TRIM / SELL / AVOID), conviction (high / medium / low), time horizon (tactical / strategic / long), an English thesis bullet list, earnings catalysts, the playbook inputs used to reach the call, any conflicts raised by the cross-cutting guards, and a drill_down pointer to the next tool. Each holding is classified into growth / quality / hedge / income / speculative and rated by the matching playbook; results are ranked by severity × conviction so TRIM/SELL/AVOID surface first. A correlation guard runs across every BUY/ADD candidate: if a starter position would lift the portfolio's weighted-average 90d pairwise correlation by more than 0.10 and push it above 0.70, the conviction is downgraded one tier and a correlation_guard conflict is appended. Claude reads the array top-down and narrates the verdict; it doesn't reorder or re-derive the calls.
• Watchlist — new-ticker candidates surfaced from peers of what you already hold, reordered within each relation so the lowest-correlation candidates come first
• Freshness — last-synced dates for prices and FX, so Claude knows when to sync first
• Portfolio — holdings with weights, daily P&L, total cost vs market value
• Concentration — HHI by ticker, sector, currency, country, plus effective_n and the 90d pairwise correlation matrix (firma sync rebuilds the cache from snapshot history)
• Movers — top winners and losers for the day
• News — recent headlines per holding (with rule-derived sentiment + mechanism attribution)
• Earnings — upcoming earnings dates with EPS estimates
• Economic calendar — high/medium-impact events for the week
• Macro — FRED snapshot (VIX, 10Y yield, credit spread, fed funds, FX impact in home currency)
• Stress & Regime — Economic Stress Index + Risk-on / Mixed / Risk-off bias
• Commodities — WTI oil, gold, copper (via FRED — no extra key)
• Dividend calendar — upcoming ex-dates and estimated income for held positions
• Goal tracking — measured CAGR vs needed CAGR (no projected net worth — forward-looking probabilities route to project_net_worth)

If today's snapshot is missing, get_market_brief records one automatically before assembling the brief, so daily history accrues even when you forget to run firma add snapshot.

Diagnostic drill-down tools: show_portfolio, show_txns, show_balance, show_flow, show_snapshot, show_concentration, show_dividend, show_news, show_financials, show_valuation, show_earnings, show_macro, show_fx, show_risk, show_benchmark, show_profile, show_thesis_track (past trades with recorded reason ↔ outcome — for thesis review and behavioral anchoring), show_anti_portfolio (post-exit performance of fully-sold positions — surfaces missed_rebound vs good_call labels with summary counts).

MCP Resources (attach as context with @firma:…): firma://portfolio (current holdings table), firma://thesis-log (recent trades with reasons), firma://profile (your goals). Use these in Claude Code when you want the model to reason about your portfolio without burning a tool call.

Forward projection

project_net_worth is the only thought-experiment tool today. Unlike a generic Monte Carlo calculator, it operates on your actual portfolio and treats assumption disclosure as a hard contract.

• Multi-scenario in one call. Pass up to 8 scenarios (e.g. bear / base / bull, or "historical vs stressed"). Each scenario has its own label + expected_annual_return + annual_volatility. All scenarios share the same seeded shock sequence so cross-scenario differences reflect (return, vol) — not random noise. Apples-to-apples.
• Any horizon, one tool. horizon_months accepts 1 (next-month outlook) through 720 (60-year FIRE planning). Same GBM math, same fan-chart output. No artificial short/long split.
• Optional target. Supply target_value for probability_of_reaching_target + median_months_to_target. Omit it for the pure distribution at the horizon.
• Caller-supplied assumptions, always disclosed. Claude is required to surface the (return, vol) it used for each scenario verbatim — so the user sees what's modeled, not a hidden "advisor view." A natural anchor scenario is the user's measured risk_summary.annualized_return_pct and annualized_vol_pct from the brief — Claude reads those and uses them as one scenario among others.
• Output. Per scenario: P10 / P25 / P50 / P75 / P90 final values, percentile trajectories (yearly resolution by default; monthly available), probability of reaching target, median months to reach, plus the (return, vol) echoed back. 1000 iterations per scenario by default (P-band ±1.5%); tunable up to 10000 for smoother fan charts.

"Will I retire on time?" — Claude pulls your target_net_worth and retirement_target_year from your profile, your current_net_worth and 90d realized vol from the brief, runs project_net_worth with three scenarios (historical-stats / bear / bull), shows the fan and the probability per scenario, and discloses every assumption. The brief never claims a future number.

Conditional what-if (simulate_scenario)

For the day-to-week horizon — "what if PCE comes in hot Thursday", "TSLA earnings miss by 10%", "S&P −2% tomorrow" — simulate_scenario is the deterministic sibling of project_net_worth. You supply the shocks (the assumption), it walks every holding via per-ticker beta (pulled from holdings[].fundamentals.beta_5y), and returns point-estimate P&L impact per holding and portfolio total. Optionally pass daily_vol_pct from the brief and each scenario also returns a one_sigma_range_usd, so a hypothetical move can be sized against your portfolio's natural daily volatility ("a 1.3σ event, not a tail"). High-impact macro events in the brief (FOMC, CPI, PCE, NFP within 7 days) auto-trigger this chain — Claude doesn't wait for you to ask.

Prompts — type / in Claude Desktop to access guided workflows. Firma ships five MCP prompts (no need to memorize — / opens autocomplete):

Prompt	What it does
import-trades	Confirms column mapping then inserts every row in a single add_txn call — handles 100s of rows in one shot
import-balance	Same flow for a net-worth spreadsheet (rows = months, columns = balance categories) → one add_balance call
import-flow	Same flow for an income/expense spreadsheet → one add_flow call
month-end	Walks through balance + cash flow entry for the period, submits in one batch
setup-profile	Captures your goals (birth year, retirement target, target net worth, risk tolerance, notes) so all future analysis is anchored to them — every field optional

Reasoning frameworks (no slash command — Claude applies them automatically based on how you phrase the question):

Framework	Triggered when…
pre-mortem	You're considering a buy or seem anchored to a single thesis — Claude generates five plausible failure modes tied to actual data, with portfolio impact in dollars
rebalance	Concentration drift or "how's my balance" questions — current vs target, specific trades, wash-sale checks
tax-harvest	Year-end or tax-related questions — loss scan, wash-sale rule, replacements, savings by bracket
scenario	"What if X happens" questions — routed to simulate_scenario with explicit beta assumptions Claude discloses

setup_status (called by Claude at session start) returns the analyst_context block — the fund-manager persona principles + data-gating rules + the four frameworks above. This is loaded on every conversation, so the right framework fires without you typing anything.

Historical FX note: firma sync fx backfills daily FX rates (KRW, JPY, EUR, CNY, GBP, HKD, INR, TWD) from FRED starting at your earliest transaction date. Subsequent runs are increment-only. This means a 2018 balance in KRW uses the 2018 rate — not today's.

Architecture

Firma is a Yarn Berry monorepo with a layered, port-and-adapter design. The application layer never imports an external-API package directly — it talks to domain interfaces, and adapters implement those interfaces against specific services.

Workspace layout reflects the layers:

packages/
  domain/           ports + types + routers, no external-API dep
    domain/         market/ (Quote, StockSnapshot, Candle, news, …),
                    macro/ (series, country indicators, live FX, stress,
                    regime), event/ (world disaster events)
  external-api/     raw API clients, no firma-domain dep
    finnhub/  fred/  worldbank/  gdacs/  open-er-api/  yahoo-finance/
  adapter/          the only layer that imports both sides
    adapters/       provider registry — createMarketProviders, …
  use-case/         business logic
    portfolio/      brief/   advisor/
  shared/           cross-cutting
    db/             utils/
apps/
  cli/              mcp/   docs/

Import dependencies (A → B means "A imports B"):

apps/cli, apps/mcp
  → @firma/adapters        data-clients.ts builds providers via the registry
  → @firma/domain          commands/tools call domain clients directly
                           for simple reads (e.g. `show news`)
  → use-case packages      richer flows (portfolio sync, brief)

use-case (portfolio, brief, advisor)
  → @firma/domain          brief also composes @firma/advisor

@firma/adapters
  → @firma/domain          implements MarketDataProvider / MacroDataProvider /
                           EventDataProvider
  → external-api packages  calls the raw HTTP/RSS clients

external-api (finnhub, fred, worldbank, gdacs, open-er-api, yahoo-finance)
  → shared only            pure API clients; no domain/adapter knowledge

@firma/domain
  → shared only            the stable core; no external-api / adapter /
                           use-case / app dependency

shared (db, utils)
  → nothing                cross-cutting leaf; imported by every layer above

At runtime, data flows the opposite way:

External API → @firma/adapters → domain client → use-case → cli / mcp

Dependency rule: external-API packages have zero firma-domain knowledge; domain packages have zero external-API knowledge. @firma/adapters is the only package that imports both sides. Application code (apps/cli, apps/mcp) talks to domain clients; the one place it touches @firma/adapters is data-clients.ts, which constructs the concrete providers.

Capabilities — each domain interface declares a capability set; the router picks a provider per capability:

Sub-domain	Capabilities
market	quote, snapshot, candles, news, insider, financials, earnings, dividends, peers, economic-calendar
macro	series, metadata, search, country-indicator, live-fx
event	disasters

Why this matters: adding a data source touches no command or business-logic code. Adding KRX coverage, for example, means a new packages/external-api/naver/ raw client, a createNaverProvider(): MarketDataProvider in packages/adapter/adapters/src/, and one registration line in packages/adapter/adapters/src/registry.ts. See CONTRIBUTING.md for the full recipe.

Development

Requires Node.js 22+ and Yarn Berry.

corepack enable
yarn install

yarn dev:cli show portfolio    # CLI dev mode
yarn typecheck                 # Full type check
yarn test                      # Run unit tests

See CONTRIBUTING.md for the full architecture walkthrough and the recipe for adding a new external data provider.

Works with Herald

Firma is the financial intelligence layer of the Herald ambient voice assistant stack. When connected, Herald can:

• report your portfolio performance and net worth by voice
• answer "how much did I spend on fees this quarter?" without opening a spreadsheet
• surface market context mid-conversation alongside your own position data

Herald + Firma + Memex — ambient voice, financial intelligence, and persistent memory in one personal AI stack.

llms.txt

llms.txt is a machine-readable summary of this project for LLM agents — concise description with documentation links, following the llms.txt standard.

License

MIT © Evan Moon