Tragent researches US equity strategies with LLM agents and trades the accepted ones through Interactive Brokers. It supports multiple users, each with their own IBKR account.

Trading currently runs on IBKR paper accounts only. Live account support is coming soon.

_{Screenshots use demo data.}

Research:

• A Research Agent and a Critic Agent, each a LangChain create_agent loop with its own tools. The Critic can accept a draft, reject it, or send it back for one revision.
• Requests are written in plain language. The parsed goal is shown for confirmation, and missing details are asked for, before any research starts.
• Market data, backtests, costs, and metrics are computed in Python. The model only reads compact summaries of the results.
• Out-of-sample results stay hidden until a candidate is locked. Agent steps and tool calls have fixed budgets.
• SEC filings, Yahoo Finance news, StockTwits, and Reddit posts are retrieved through pgvector as supporting context.
• Tasks run on Celery workers and are checkpointed in PostgreSQL after every step, so they keep running after the browser closes and can be resumed after a failure.

Trading:

• Each linked IBKR login runs in its own IB Gateway container.
• Orders are idempotent and sent to a single broker process over Kafka. The broker reconciles with the Gateway and resends any lost command.
• Orders, balances, and positions update in the browser over a WebSocket.
• Deployed strategies run every trading day at 09:00 ET with the same code as the backtest.
• IBKR Activity Statements (English or Chinese) can be imported, and the trade history can be exported as CSV.

flowchart LR
    B[Browser] --> N[Nginx]
    N --> W[Vue app]
    N --> A[FastAPI]
    N --> S[(SeaweedFS)]
    A --> P[(PostgreSQL)]
    A --> R[(Redis)]
    R --> C[Celery worker]
    BT[Celery beat] --> R
    C --> L[LLM]
    C --> P
    C --> S
    A <--> K[(Kafka)]
    K <--> BR[Broker]
    BR --> G[IB Gateway per user] --> I[IBKR]

All strategies are long-only.

Minute strategies close out every position by the end of the session. Yahoo Finance only keeps 30 days of minute bars, so a minute study covers at most 20 sessions.

Requirements:

• Docker with Compose
• An OpenAI-compatible chat model and an embeddings endpoint
• A LangSmith API key
• An IBKR login for each user who links an account
• Python 3.13+ and Node.js 22.18+ for local development

Copy the environment file, fill it in (see the comments in .env.example), and pull the Gateway image:

cp .env.example .env
docker pull ghcr.io/gnzsnz/ib-gateway:latest

docker compose --profile app up -d --build

Open http://localhost:8080, register, and link an IBKR login on the Account page. Migrations run before the app starts.

Start the infrastructure, install the package, and migrate:

docker compose up -d
python -m pip install -e ".[dev]"
alembic upgrade head

Then start each process in its own terminal:

uvicorn tragent.api.app:create_app --factory --reload

celery -A tragent.worker worker --concurrency=16 --loglevel=info

celery -A tragent.worker beat --loglevel=info

python -m tragent.trading.broker

cd frontend && npm ci && npm run dev

The frontend runs at http://localhost:5173 and proxies /api to FastAPI. API docs are at http://127.0.0.1:8000/docs. The API needs access to the Docker socket to manage Gateways, and POSTGRES_POOL_MAX_SIZE should be at least twice the worker concurrency.

To run more API replicas:

docker compose --profile app up -d --scale api=2

Workers can run on any machine that reaches PostgreSQL, Redis, Kafka, and SeaweedFS.

The frontend and API can be deployed separately on subdomains of the same root domain, such as app.example.com and api.example.com. The refresh token cookie is SameSite=Strict, so different root domains won't work.

docker build --build-arg VITE_API_BASE_URL=https://api.example.com -t tragent-web frontend

CORS_ORIGINS=["https://app.example.com"]
S3_PUBLIC_URL=https://files.example.com

python -m pytest
ruff check src tests scripts
cd frontend && npm run type-check && npm run test:unit -- --run

Tests don't call external services. Agent evaluations use a real model:

python scripts/run_agent_evals.py               # decision scenarios
python scripts/run_agent_evals.py --with-smoke  # plus an end-to-end run

src/tragent/
├── api/          # FastAPI app, auth, diagnostics
├── research/     # research workflow, persistence, reports
│   └── agent/    # agents, tools, prompts
├── quant/        # market data, strategies, backtests
├── trading/      # Gateways, broker, orders, deployments, statements
├── storage.py    # object store
└── worker.py     # Celery app and schedule
frontend/         # Vue app and Nginx config
migrations/       # Alembic migrations
scripts/          # agent evaluations

MIT

Backtest results are not investment advice.