How It Works

Annie's Gold Tracker started from a simple question: if a feasibility study assumed gold at US$1,800/oz and gold is now US$4,200/oz, how much is the project actually worth? That question turned into a pipeline that scrapes ASX announcements, extracts economics from PDFs with AI, and revalues every project at today's gold price. Built using Claude Code (Anthropic's AI coding assistant). Still early, lots to improve.

AI and tools used in this project:

• Claude Code (Opus 4.6): AI pair-programming assistant that wrote the application code, designed the architecture, and built the data pipeline. Every file in this repo was authored through conversation with Claude Code.
• Claude Haiku 4.5 API: Reads feasibility study PDFs and extracts structured financial data (NPV, IRR, AISC, reserves, production, mine life, gold price assumption) via the Anthropic Messages API. Cost is about $0.01 per study.
• Claude Sonnet 4.6 API: Fallback model for PDFs that exceed Haiku's 200K token context window.
• Custom MCP Server (mcp-asx-gold): A Model Context Protocol server that gives Claude Code access to live ASX data. Provides 9 tools: stock quotes, batch quotes, historical prices, company financials, announcements, gold price, gold-vs-stock comparison, stock returns, and a pre-built developer list.
• Yahoo Finance API: Stock prices for all 30 tracked companies, gold futures (GC=F), and AUD/USD exchange rates. No API key required.
• ASX MarkitDigital API: Company announcements, feasibility study PDF downloads, and market cap data (via the /header endpoint) from asx.com.au. No API key required. Market cap was originally sourced from Yahoo Finance but it returned null for all ASX stocks, so this was switched to the ASX API after a data audit.

Before a gold mine gets built, the company publishes a feasibility study. This study models the project's economics (how much it will cost, how much gold it will produce, and how profitable it will be) all based on a specific gold price assumption. For example, a study might assume gold at US$1,800/oz.

The key insight:these studies are snapshots in time. The gold price used in the study is whatever gold was trading at when the study was written. But gold prices change. If gold has risen from US$1,800/oz to US$4,200/oz since the study was published, then the project's actual economics are much better than what the study says. The NPV (Net Present Value, basically what the project is worth) could be double or triple the published number.

Why this matters: the stock market is slow to reprice small companies. Big miners with analyst coverage get repriced quickly, but a small ASX developer with a $50M market cap might still be valued based on a study that assumed gold at US$1,800/oz, even though gold is now US$4,200/oz. That gap between what the study says and what current gold prices imply is the opportunity this tool looks for.

The tool automates this process: it scrapes ASX announcements for new feasibility studies, extracts the key numbers (including what gold price the study assumed and when it was published), and recalculates the NPV at today's gold price.

The pipeline runs twice daily at 7am and 7pm AEST (Australian market hours). It executes four stages in sequence, each building on the last.

The engine uses a revenue-sensitivity approach rather than rebuilding the full DCF model. This is deliberately conservative: it captures the marginal impact of gold price changes without speculating on operating leverage, currency effects, or development risk.

All data flows through a SQLite database with full audit trail. Every extraction and revaluation is logged with the model used, prompt sent, response received, and timestamp. This makes every number on the dashboard traceable back to a specific page in a specific PDF.

• Linear approximation. The revenue-sensitivity approach assumes NPV scales linearly with gold price for marginal changes. This breaks down for very large price moves where operating leverage, grade cutoff optimisation, and mine plan changes would materially alter the economics.
• No operating leverage.Higher gold prices can justify mining lower-grade ore, extending mine life and increasing total production. The engine doesn't capture this upside.
• Currency effects. Many ASX studies report in AUD. Gold trades in USD. The FX conversion uses the current AUD/USD rate, which may differ from the rate assumed in the study.
• Development risk.A developer's market cap rightly trades at a discount to NPV. Construction risk, permitting risk, financing risk, and execution risk all justify a gap between NPV and market cap. A large NPV uplift does not mean the stock should trade at NPV.
• ASX API coverage. The MarkitDigital API only returns recent announcements (~20-50). Older feasibility studies require manual upload or company website scraping.
• AI extraction accuracy. Claude Haiku performs well on well-structured feasibility study PDFs, but investor presentations with charts-only data or non-standard layouts may produce null values. All extractions include source page numbers so results can be manually verified against the original PDF.
• Removed metrics.Earlier versions displayed a "confidence score" (AI self-assessment, always ~95%), a "mispricing score" (broken when market cap data was missing), and a "gold leverage score" (non-standard metric). These were removed for transparency.

The mcp-asx-gold server is a standalone Node.js package in the mcp-asx-gold/ directory. It uses the official @modelcontextprotocol/sdk and communicates over stdio transport. Claude Code calls these tools during development to get live data.

server.tool(
  "asx_stock_quote",
  "Get live stock price for any ASX-listed company",
  { ticker: z.string().describe("ASX ticker symbol") },
  async ({ ticker }) => {
    const quote = await getStockQuote(ticker);
    return { content: [{ type: "text", text: JSON.stringify(quote) }] };
  }
);

{
  "ticker": "TCG",
  "name": "TURACOGOLD FPO",
  "price": 0.545,
  "currency": "AUD",
  "marketCap": 574.7,    // A$M, from ASX /header endpoint
  "volume": 2143188,
  "fiftyTwoWeekHigh": 0.89,
  "fiftyTwoWeekLow": 0.40,
  "exchange": "ASX",
  "fetchedAt": "2026-06-19T06:02:52Z"
}

Every action in the system is logged. The goal is that any number on the dashboard can be traced back to its source: which PDF, which page, which model, what prompt was sent, what came back, and when.

audit_log {
  id            INTEGER PRIMARY KEY
  action        TEXT     -- "extraction_created", "revaluation_created", "price_updated"
  entity_type   TEXT     -- "extraction", "revaluation", "company"
  entity_id     INTEGER  -- FK to the entity that changed
  previous_value TEXT    -- JSON snapshot before change
  new_value     TEXT     -- JSON snapshot after change (full extraction JSON)
  source        TEXT     -- "ai", "system", "user"
  created_at    TEXT     -- ISO timestamp
}

extractions {
  extraction_model  TEXT  -- "claude-haiku-4-5-20251001"
  source_pages      TEXT  -- "[4, 4, 3]" (which pages the numbers came from)
  raw_extracted_text TEXT  -- full JSON response from Claude
  extracted_at      TEXT  -- when the extraction ran
}

-- Every extraction links to:
--   study (which PDF) → company (which ticker)
--   audit_log (full before/after snapshot)
--   journal (human-readable log entry)

Current state: 2 extractions out of 21 developers. Here is what would move this from prototype to useful research tool, roughly ordered by impact.

This project was built through conversation with Claude Code. Here is what that process actually looks like:

• Start with the thesis, not the code. I explained what gold feasibility studies are, why stale gold price assumptions create a valuation gap, and what data I'd need. Claude Code designed the architecture from that.
• Build the MCP server first. Before building the dashboard, I had Claude Code build the MCP server so it could access live ASX data while writing the rest of the code. This means the code was informed by real data shapes, not guesses.
• Audit ruthlessly. After the first version was built, I went through every number on the dashboard and asked "is this real or fake?" That audit found the confidence scores (AI flattering itself), the mispricing scores (broken math), and the market cap bug (Yahoo returning null). I told Claude to fix or remove everything that wasn't honest.
• Document everything. Every time we found a bug or made a decision, it went into the research journal. This creates a trail that shows the thinking, not just the output.
• Iterate on data quality before UI polish. No point making charts pretty if the underlying data is wrong. Fix the pipeline first, then the presentation.

To keep improving this dashboard, I'd continue the same loop: run the pipeline, look at what the data actually shows, ask whether each number is real or misleading, fix what's broken, and document the decisions. AI is a tool for writing the code faster, but the human judgment about what to build and what to cut is what makes the output trustworthy.