I built TradingSpy: local, privacy-first AI trading assistant(First Open Source)

Local-first AI trading research: market heatmaps, news catalysts, strategy generation, Backtrader backtests, and transparent agent runs in one Docker app.

TradingSpy is an open-source research workstation for traders and builders who want to ask questions, inspect market context, generate strategy ideas, and test them against real historical candles without wiring together five separate tools.

It is not a broker and it does not place trades. It is a local research environment for analysis, backtesting, and strategy iteration. Fully open-source, zero data privacy concerns, and free of charge.

Trading Companion— Chat with your market data, strategies, news, heatmaps, and backtest history.** Strategy Researcher**— Research to find the best trading strategies until it beats the baseline strategy.** Trading Trend Prediction**— Leverage calculation and LLM with the support and resistance lines to simulate expected stock trend.** Trading Signal Analysis**— The tool analyzes the real-time movement, incorporates peer stocks, insider trading information, and trading indicators.

TradingSpy is designed with a hybrid approach: traditional data visualisation for quick, deterministic results combined with loop-engineering-powered agents as a trading companion.

Feature	What it does
Market Intelligence
Real-time quotes, sector heatmaps, industry performance, insider activity, news search, and fundamentals — all in one query.
AI Strategy Generation
Describe a trading thesis in plain English; get a working Backtrader strategy with syntax and runtime validation.
Automated Backtesting
Every generated strategy is backtested against downloaded candles with configurable parameter sweeps.
Benchmark Comparison
Every result is compared to buy-and-hold and any saved strategy. Underperformers are rejected automatically.
Loop Engineering
Set a goal ("beat buy-and-hold", "find undervalued semiconductors") and the agent iterates until it succeeds — no babysitting required.
Transparent Agent Runs
Every tool call, validation failure, rejection reason, and accepted result is logged and visible in the Task Center.
Multi-Provider LLM
Google AI Studio, Mistral, OpenRouter, NVIDIA, LiteLLM, Ollama (local), AWS Bedrock, GCP Vertex AI, and Azure OpenAI.
OpenAI-Compatible API
Use TradingSpy as a backend for scripts, other agents, or custom integrations via /v1/chat/completions .
Local-First Architecture
All data stored under backend/data/ . No external accounts, no telemetry, no cloud dependency.

Agent	Example Prompt	What it does
Strategy Race
Generate until it beats buy and hold for QQQ. Use daily candles. Improve EMA_Trend for TQQQ using daily candles. Generate until it beats EMA_Trend, not buy and hold.
Generates strategies based on selected modes (tick data, research papers, etc.) from the AI Strategy Studio. Improves or compares strategies over rounds — can use a previous accepted version or selected baseline, generate candidates, backtest them, and accept only versions that beat the target benchmark.
Signal Analysis
Predict the next move for btc-usd for daily interval
Reads recent bars and support/resistance levels to predict the price trend.
Stock Screening
Scan AI stocks until you find 10 which are good enough on fundamentals
Uses the fundamental scanner to search for undervalued stocks. Screens a universe for valuation/growth/profitability candidates, enriches passing names with market context, news, options, and insider summaries. Can continue with a wider universe.
Chat
Give me a daily market brief with breadth, strongest and weakest industries, important news, and earnings.
Pulls data from yfinance to summarize daily market information.

If the request involves long-running work, the UI creates a background run through /api/agent/runs

. Background runs are stored locally, visible in the Task Center, and support:

Method	Endpoint	Purpose
GET
/api/agent/runs
List recent runs
GET
/api/agent/runs/{run_id}
Poll full state
POST
/api/agent/runs/{run_id}/stop
Request cancellation
POST
/api/agent/runs/{run_id}/continue
Continue a completed or stopped run
DELETE
/api/agent/runs/{run_id}
Delete a single run
DELETE
/api/agent/runs
Clean all records

For strategy workflows, the agent is deliberately conservative: it validates generated code before backtesting, rejects zero-trade results instead of treating 0% ROI

as meaningful, and reports validation failures and runtime errors as part of the public run log. It supports custom agent instructions, answer budget, run detail, sequential/parallel execution, and custom battle parameters.

For insider buy/sell questions, the assistant uses deterministic tool-backed responses. It reports only returned records, separates open-market buys/sells from grants or awards, and says so if the feed is unavailable instead of filling gaps from memory.

Not every question needs an agent. TradingSpy ships a full market dashboard for quick, deterministic results.

Component	Details
Sector Heatmap
Color-coded grid of 25+ industry proxy ETFs grouped by sector. 16 time periods (1 min – max + YTD), extended hours toggle, search/filter, custom groups, and an Explain button that sends the heatmap to the AI assistant for analysis. Two display modes: industry ETFs or watchlist stocks.
Indices Banner
Top-of-page bar showing S&P 500, Dow Jones, NASDAQ 100, and Russell 2000 with live prices and percentage changes.
Industry Movements
Tracks individual stock price changes across 12 time windows (1 min to 1 year) for 68+ major US stocks. Universe presets: High Cap, Semis, Software/AI, Leverage.
Watchlist & Intelligence
Auto-sync watchlists, real-time batch quotes, deep-dive panel (company info, technicals, news, insider activity), and embedded candlestick charts.

Source	What it provides
Yahoo Finance
Price quotes, OHLCV candles (daily, intraday, extended-hours), fundamentals, insider transactions, analyst recommendations, earnings dates, options chains, sector/industry metadata, screener queries. Primary data backbone.
SearXNG
Privacy-respecting metasearch for web and news — financial news, analyst opinions, macro events, catalyst research. Runs locally via Docker or standalone.
DuckDuckGo
Fallback web search when SearXNG is unavailable. HTML scraping + instant answer API.
arXiv
Academic papers on quantitative finance and algorithmic trading. Abstract and full-text PDF reading.
Backtrader
Local backtesting engine for strategy execution, parameter optimization, and benchmark comparison.

Any Yahoo Finance-compatible symbol works. Coverage varies by symbol and upstream source.

Market	Examples	Suffix
United States	AAPL , NVDA , QQQ , SPY
—
London	AZN.L , HSBA.L
.L
Hong Kong	0700.HK
.HK
Japan	7203.T
.T
India	RELIANCE.NS
.NS
Canada	SHOP.TO
.TO
Australia	BHP.AX
.AX
Germany / France / UK / Eurozone	^GDAXI , ^FCHI , ^FTSE , ^STOXX50E
^ prefix
China	000001.SS
.SS
Crypto	BTC-USD , ETH-USD
-USD
Commodities	GC=F (Gold), CL=F (Oil)
=F

Region	Indices
United States	S&P 500, Dow Jones, NASDAQ 100, Russell 2000, VIX
Europe	STOXX 50, FTSE 100, DAX, CAC 40
Asia	Nikkei 225, Hang Seng, Shanghai Composite, ASX 200
Commodities	Gold Futures, Crude Oil
Crypto	Bitcoin, Ethereum

Provider	Environment variable	Example default model
Google AI Studio	GOOGLE_AI_STUDIO_API_KEY
gemini-2.5-flash
Mistral	MISTRAL_API_KEY
mistral-large-latest
OpenRouter	OPENROUTER_API_KEY
openai/gpt-4o-mini
NVIDIA	NVIDIA_API_KEY
nvidia/llama-3.1-405b-instruct
LiteLLM	LITELLM_API_KEY , LITELLM_BASE_URL
Your proxy's model ID
Ollama (local)	OLLAMA_BASE_URL ; no API key required
qwen2.5-coder:7b

Additional providers: AWS Bedrock, GCP Vertex AI, and Azure OpenAI are supported via the LiteLLM proxy. PointLITELLM_BASE_URL

at your proxy and configure provider credentials there.

Keys may be stored in .env

/backend/.env

or entered in the app's Settings page. Never commit a real key. See .env.example for every supported setting.

git clone https://github.com/mrhustlex/TradingSpy-TradingAgentService.git
cd TradingSpy
cp .env.example .env

Add at least one provider key to .env

:

GOOGLE_AI_STUDIO_API_KEY=your-gemini-key
DEFAULT_PROVIDER=google_ai_studio
DEFAULT_MODEL=gemini-2.5-flash

Or use Ollama (no API key required):

ollama pull qwen2.5-coder:7b
DEFAULT_PROVIDER=ollama
DEFAULT_MODEL=qwen2.5-coder:7b
OLLAMA_BASE_URL=http://host.docker.internal:11434/v1

You can also configure providers later in the app's Settings page.

docker compose up -d --build

Service	URL
App

http://localhost:8000http://localhost:8000/docshttp://localhost:8080

docker compose down

Runtime data remains under backend/data/

. Pull updates and rebuild with git pull && docker compose up -d --build

.

cd backend
python3.11 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
uvicorn main:app --reload --host 0.0.0.0 --port 8000

Use Python 3.11. The pinned data-science dependencies are not reliable with Python 3.13.

cd frontend
npm ci
npm run dev

Open http://localhost:5173.

npm run dev:searxng    # start
npm run stop:searxng   # stop

This starts only SearXNG at localhost:8080

. Alternatively, docker compose up -d searxng

.

flowchart LR
    User["User / Browser"] --> Frontend["React Frontend<br/>localhost:3000"]
    Frontend --> Backend["FastAPI Backend<br/>localhost:8000"]
    Backend --> ChatAgent["Tool-Using Chat Assistant<br/>short research + tool checks"]
    Backend --> WorkflowAgent["Background Workflow Agents<br/>strategy_create / strategy_race / market_review / fundamental_screener"]
    Backend --> RemoteAgents["Remote Agent Outputs<br/>OpenAI-compatible / ACP / A2A"]
    Backend --> Backtest["Backtrader Engine<br/>backtests + optimization"]
    Backend --> Market["Market Intelligence<br/>yfinance + heatmaps + news"]
    Backend --> Store["Local Data<br/>TinyDB + candles + strategies"]
    Backend --> Search["SearXNG<br/>localhost:8080"]
    ChatAgent --> LLM["Validated LLM Providers<br/>Google AI Studio / Mistral / OpenRouter / LiteLLM"]
    WorkflowAgent --> LLM
    RemoteAgents --> ChatAgent
    RemoteAgents --> WorkflowAgent

All runtime data is stored locally and ignored by Git:

backend/data/
├── db.json
├── system_settings.json
├── market_data/local_user/
├── strategies/local_user/
├── results/local_user/
├── optimization_history/
└── temp_datas/

Back these up separately if the results matter to you.

What	Deterministic?
Saved strategy against same candles, dates, capital, commission, parameters	Yes
LLM-generated strategy code	No — non-deterministic across runs
Live quotes, fundamentals, insider records, heatmaps, news	No — changes over time
Model aliases and upstream provider behavior	May change — use explicit model IDs when comparing
Backtest performance	Depends on period and assumptions — not a promise of future returns

Keep the dataset, generated strategy, benchmark, and run details together when sharing a result.

Concern	Detail
Research only	TradingSpy is for research and education. It is not financial advice.
Backtest overfitting	Backtests can overfit and do not predict future returns.
Code execution	Generated strategy Python is executed locally and is not sandboxed. Review it before running.
Network binding	Keep all services bound to localhost unless you add auth, TLS, network controls, and process isolation.
Credentials	Keep API keys out of git. Use .env or your own secret manager.

Task	Command
Check services	docker compose ps
Health check	curl http://localhost:8000/health
View backend logs	docker compose logs -f backend
View frontend logs	docker compose logs -f frontend
View SearXNG logs	docker compose logs -f searxng
Full rebuild	docker compose build --no-cache && docker compose up -d
Check disk usage	docker system df
Prune build cache	docker builder prune

• Discord community server
• Per-agent GIF demos in README
• More LLM provider integrations
• Strategy sharing and export

Join the conversation:

[Discord](coming soon)

Contributions are welcome. See CONTRIBUTING.md for the full guide.

Type	What to do
Bug reports	Open an issue with steps to reproduce, expected vs. actual behavior, and environment details.
Feature requests	Describe the use case, not just the implementation. What problem does it solve?
Code	Pick an open issue or start a discussion first for large changes.
Documentation	Fix typos, clarify explanations, or add examples.
Testing	Try edge cases, different providers, or non-US markets and report what breaks.

• Fork the repository and create a feature branch.
• Set up with Docker or follow Manual Development. - Run development checks:

python3 -m py_compile backend/main.py backend/modules/*.py
npm run build --prefix frontend
npm run lint --prefix frontend

• Add or update tests where practical.
• Include screenshots for visual changes and request/response examples for API changes.
• Submit a pull request with a clear description of what changed and why.

Rule	Why
One logical change per PR	Keeps reviews focused and diffs clean.
Never commit credentials, databases, market data, strategies, caches, or build output	Preserves security and keeps the repo small.
Treat generated strategy code as untrusted	Preserve the local-only security model.
No new lint warnings or errors in files you change	Keeps the codebase healthy.
Contributions licensed under the repository's license	Standard open-source contribution terms.

TradingSpy is licensed under the PolyForm Noncommercial License 1.0.0. Non-commercial use is allowed; commercial use requires separate permission from the copyright holder. See LICENSE.

TradingSpy is experimental software. It is not investment advice, a trading signal service, or a guarantee of performance. You are responsible for reviewing all generated code, assumptions, data quality, and results.

source & further reading

github.com — original article