X Copy-Trade Tool
A cautious pipeline that watches specific X accounts for trade calls, has an LLM turn them into a strict structured signal, independently validates and risk-checks that signal, and — depending on mode — logs it, proposes it for human approval, or submits it to a real OANDA practice account. Never connected to live money.
What it does
Every post that comes in goes through the same pipeline, end to end, before anything is ever proposed or submitted:
Architecture
Posts arrive through one of three input adapters (manual/JSON, a webhook fed by a companion browser extension, or the full X API v2) and land in a SQLite audit trail before anything else touches them. A rule-based pre-filter plus an OpenAI structured-output call classifies each post; a context engine pulls in prior posts from the same author only when it can state an explicit reason two posts are related, rather than assuming adjacency means continuity. A second OpenAI call extracts a strict structured signal, with hard-coded (not just prompted) safeguards — every evidence fragment must appear verbatim in the source post. That signal then passes through deterministic validation (staleness, instrument mapping, stop/target logic, market hours, conflicting positions) and a separate risk manager with its own numeric thresholds and a circuit breaker, before reaching one of three approval modes and, finally, an OANDA v20 practice-only broker integration that structurally refuses to construct against the live endpoint.
The dashboard itself is a server-rendered FastAPI + Jinja2 page with vanilla JS polling — deliberately no frontend build step. Public visitors get full read access to every tab; every write (approve/reject, kill switch, tracked accounts, manual posts) requires an operator session, enforced server-side by middleware rather than just a disabled button.
Design principles
- Practice only, structurally. The broker layer refuses to construct against OANDA's live endpoint at all — this isn't a config flag away from real money.
- Evidence must be verbatim. The signal extractor's cited evidence is checked against the actual source post text, not trusted from the model.
- Validation and risk are independent. Deterministic checks (staleness, instrument mapping, stop/target logic) run separately from the risk manager's numeric thresholds — one bad number can't silently waive the other.
- Everything is recorded. Every post, classification, signal, proposal, order, and trade is written to an audit trail, whether or not it ever became a trade.
- Public means read-only, by default. Anyone can watch it work. Only a logged-in operator can change anything.
Tech stack
Open source
The whole thing is open source — classification/extraction, risk management, the OANDA integration, this dashboard, and this landing page. Clone it and run your own copy against your own OANDA practice account in a few minutes:
git clone https://github.com/1234Es/x-copy-trade-tool-demo.git
cd x-copy-trade-tool-demo
python -m venv .venv && source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env # fill in OPENAI_API_KEY at minimum
python -m app.main # open http://127.0.0.1:8000That gets the dashboard running locally with manual/webhook post input. Full setup (OANDA practice credentials, the X API adapter, the browser extension, a security checklist before leaving it running) is in the repo's README.