Corporate insider purchases and congressional trade disclosures become ranked signals, market context, AI allocations, demo orders, and an inspectable performance record.
Agent Trading is an autonomous demo-account research experiment built around a simple question:
Can public buying disclosures become a repeatable, measurable investment signal when they are scored consistently and evaluated with broader market context?
Twice a week, the system looks for recent purchases disclosed by corporate insiders and members of the US Congress. It ranks those signals, enriches the strongest candidates with fundamentals, technical indicators, news, earnings, and trading history, then asks Claude to allocate a capped demo budget.
Orders are sent only to a Trading 212 practice account. Every run is reported and the portfolio is published to a dashboard so the experiment can be evaluated instead of remembered selectively.
OpenInsider surfaces SEC Form 4 open-market purchases by executives, directors, and other company insiders.
The system gives more weight to:
- recent purchases
- larger increases in the buyer's ownership
- C-suite buyers such as a CEO, CFO, COO, president, chair, or CTO
- several insiders buying the same company
- large individual purchases
OpenInsider is an intermediary for SEC-derived data; this project does not scrape SEC filings directly.
Capitol Trades surfaces public House and Senate trading disclosures.
The system:
- estimates transaction value from the disclosed range midpoint
- favors more recent transactions
- groups multiple disclosures by ticker
- filters pure congressional candidates above the configured market-cap ceiling
- merges a ticker when both signal sources identify the same company
Congressional reporting can be delayed, so these disclosures are treated as noisy experimental signals rather than timely privileged information.
flowchart LR
A["OpenInsider purchases"] --> C["Score and group by ticker"]
B["Congressional disclosures"] --> C
C --> D["Merge overlapping signals"]
D --> E["Add fundamentals, technicals, news, earnings, and history"]
E --> F["Claude portfolio decision"]
F --> G["Budget and exposure checks"]
G --> H["Trading 212 demo orders"]
H --> I["Report and public dashboard"]
Corporate insider transactions use ownership change, buyer seniority, and recency. Congressional transactions use estimated disclosed value and recency.
Each candidate is enriched with:
- company fundamentals and market capitalization
- 1-month, 6-month, and 1-year price returns
- RSI, MACD, Bollinger Bands, and moving averages
- upcoming earnings information
- recent headlines
- historical insider-buy activity
The active pipeline has one AI decision stage. Claude receives the enriched candidate set, current demo positions, and the budget for the run. It returns ranked buy picks, allocation percentages, reasoning, optional sell recommendations, confidence, and a short market summary.
Before execution, the system enforces practical limits:
- demo account only
- configurable budget per run
- configurable maximum portfolio investment
- maximum number of picks
- existing-position exposure checks
- a short recently-traded blacklist
- fallback to the next candidate when an order cannot be placed
The GitHub Pages dashboard makes the experiment observable:
- current demo holdings and position-level returns
- invested capital, portfolio value, and unrealized P&L
- portfolio history over time
- comparison with the S&P 100
- recent decision runs and selected stocks
- the model's reasoning and confidence
The committed dashboard data is automatically refreshed and will change over time. The snapshot dated June 12, 2026 contained:
| Signal | Value |
|---|---|
| Recorded decision runs | 21 |
| Portfolio history points | 70 |
| Open demo positions | 89 |
| Demo capital invested | €46,220.33 |
| Demo portfolio value | €48,348.67 |
| Unrealized demo return | +4.60% |
This is a point-in-time demo result, not evidence of future returns or a validated investment strategy.
All times use Europe/Berlin by default and can be changed through environment
variables.
| Time | Days | Activity |
|---|---|---|
10:00 |
Monday-Friday | Refresh portfolio prices and dashboard |
15:30 |
Monday-Friday | Refresh portfolio prices and dashboard |
17:10 |
Tuesday and Friday | Build signals, ask Claude, and place demo orders |
17:35 |
Tuesday and Friday | Capture end-of-day state, report, and publish dashboard |
| Area | Responsibility |
|---|---|
| 🧾 Market-data tools | Public disclosures, prices, fundamentals, technicals, earnings, and news |
| 🧠 Agent pipeline | Structured Claude decision from the fully enriched digest |
| 🎛️ Supervisor | Signal merging, filtering, candidate selection, and cycle control |
| 💶 Trade executor | Budget allocation, exposure checks, and order fallback |
| ⏱️ Scheduler | Trade cycles, reports, and dashboard refreshes |
| 📣 Notifications | Optional Telegram updates |
| 📉 Reporting | Markdown reports, terminal summaries, and GitHub Pages data |
agent-trading/
├── src/
│ ├── agents/ Claude decision stage
│ ├── mcp_servers/ market-data and demo-broker tools
│ ├── orchestrator/ signal pipeline, controls, scheduling, and execution
│ ├── notifications/ optional Telegram messages
│ ├── reporting/ reports and dashboard generation
│ └── config.py environment-driven settings
├── scripts/ scheduler, one-off run, dry run, report, schedule check
├── docs/ public dashboard and changing data snapshot
└── tests/ strategy plumbing and integration-contract tests
1. Install the environment
git clone https://github.com/thecr7guy2/agent-trading.git
cd agent-trading
uv sync2. Configure the demo services
cp .env.example .envRequired values:
ANTHROPIC_API_KEY=...
T212_API_KEY=... # Trading 212 practice-account key
T212_API_SECRET=...News, fundamentals fallback, Telegram, budget, portfolio cap, signal lookbacks, and
schedule settings are optional or have defaults. See .env.example.
3. Run without placing orders
uv run python scripts/dry_run.py
uv run python scripts/dry_run.py --budget 1500 --lookback 7This runs signal collection, enrichment, and Claude selection but does not submit broker orders.
4. Inspect or operate the scheduler
uv run python scripts/check_schedule.py
uv run python scripts/report.py
uv run python scripts/run_scheduler.pyscripts/run_daily.py can place demo orders. Read the code and configuration before
using that command.
- All inputs are public disclosures and public market data.
- OpenInsider and Capitol Trades are third-party sources whose availability can change.
- Corporate and congressional reports may arrive after the market has reacted.
- Claude output is non-deterministic and can vary between runs.
- The current strategy is primarily buy-oriented; exit logic is less developed.
- A benchmark comparison does not control for risk, timing, concentration, or selection bias.
- Demo fills and execution conditions may differ from real-money trading.
This repository is an educational software experiment, not financial advice or an invitation to trade. It is configured for a demo account and should not be connected to real capital without independent review, testing, risk controls, and professional financial guidance.