Deploy and Host TradingAgents on Railway

TradingAgents is an open-source multi-agent LLM framework by Tauric Research that mirrors the dynamics of a real trading firm — analysts, researchers, traders, and risk managers debate and converge on a buy/hold/sell recommendation for any ticker. Self-host TradingAgents on Railway to run the full agent graph through a clean web UI and REST API, without exposing your LLM API keys to a third-party service.

This Railway template wraps the upstream Python package in a FastAPI service so you can submit (ticker, date) analyses from a browser, poll job status, and inspect decisions through /docs. It bundles a persistent volume for analysis logs, agent memory, and market-data cache so multi-day research stays available across redeploys.

Getting Started with TradingAgents on Railway

Once the deploy finishes, open the generated *.up.railway.app URL — the home page is a single form for ticker, date, LLM provider, and debate rounds. Pick OPENAI (the default), submit, and the job will run in the background; the page polls every three seconds until the agent graph returns a decision. Visit /docs to use the auto-generated Swagger UI directly, or POST /api/analyze from any HTTP client. All historic jobs persist in /data/logs on the mounted volume.

About Hosting TradingAgents

TradingAgents orchestrates seven specialized LLM agents through a LangGraph state machine: fundamentals analyst, sentiment analyst, news analyst, technical analyst, bull researcher, bear researcher, and a trader who weighs all four reports against a risk manager. Each agent has its own prompt, tool access, and memory.

Key features:

• Multi-provider LLM support (OpenAI, Anthropic, Google, DeepSeek, Qwen, GLM, OpenRouter, local Ollama)
• Configurable debate depth — more rounds = more deliberation, more tokens
• yfinance built in (free); Alpha Vantage optional for premium data
• Persistent agent memory + decision log on disk
• Built-in checkpoints — resume long analyses after a redeploy

Why Deploy TradingAgents on Railway

Railway is the fastest path from git push to a publicly reachable agent endpoint:

• One-click Docker deploy from this template
• Persistent volume pre-mounted for memory, cache, and logs
• Generated public domain with HTTPS — usable from any client
• Built-in env var management for LLM API keys
• Bump CPU/RAM with a slider when debate rounds get heavy

Common Use Cases

• Run automated nightly analyses on a watchlist via cron + the REST API
• Compare LLM providers' decisions on the same ticker as a research baseline
• Power an internal Slack bot that returns multi-agent deliberation on demand
• Back-test agent decisions against historical data using the date parameter

Dependencies for TradingAgents on Railway

This template runs as a single service:

• TradingAgents — Python 3.12-slim image, installs tradingagents from github.com/TauricResearch/TradingAgents plus a FastAPI/uvicorn wrapper

Environment Variables Reference

Deployment Dependencies

• Source: https://github.com/TauricResearch/TradingAgents
• Latest release: v0.2.4
• Runtime: Python 3.12 (Debian slim base)
• License: Apache 2.0
• Volume: /data (1 GB recommended)

Hardware Requirements for Self-Hosting TradingAgents

LangChain + multi-agent debate is memory-hungry; 2 GB will OOM on long debates. Bump to 4 GB for production use.

Self-Hosting TradingAgents

The fastest path is the Railway template above. To run it locally for development, the upstream supports pip install plus a CLI:

git clone https://github.com/TauricResearch/TradingAgents
cd TradingAgents
python -m venv .venv && source .venv/bin/activate
pip install .
export OPENAI_API_KEY=sk-...
tradingagents

To call the analysis programmatically — useful when integrating with another service — import the graph directly:

from tradingagents.graph.trading_graph import TradingAgentsGraph
from tradingagents.default_config import DEFAULT_CONFIG

config = DEFAULT_CONFIG.copy()
config["llm_provider"] = "openai"
config["deep_think_llm"] = "gpt-4o"
config["quick_think_llm"] = "gpt-4o-mini"

ta = TradingAgentsGraph(config=config)
_, decision = ta.propagate("NVDA", "2026-01-15")
print(decision)

The Railway template wraps exactly this call inside a FastAPI background task.

How Much Does TradingAgents Cost to Self-Host?

The TradingAgents framework is fully open source under Apache 2.0 — no per-seat fees, no commercial tier. On Railway you only pay for compute, RAM, and the volume; with light usage that's under $10/month. The real cost is the LLM tokens: a single analysis with gpt-4o deep-think + gpt-4o-mini quick-think and one debate round runs roughly $0.20–$0.80 per ticker. Doubling debate rounds roughly doubles cost.

FAQ

What is TradingAgents and why self-host it? TradingAgents is a multi-agent LLM framework that simulates a trading firm's research workflow. Self-hosting on Railway keeps your tickers, prompts, and decisions on infrastructure you control instead of routing them through a third-party SaaS.

What does this Railway template deploy? A single Python service that installs the upstream tradingagents package, wraps it in FastAPI, and exposes a web UI plus REST API at /, /docs, and /api/analyze. A persistent volume at /data stores logs, agent memory, and the market-data cache.

Why does this template need a persistent volume? TradingAgents writes decision logs, market-data caches, and the agent memory file across runs. A Railway volume keeps them across redeploys so the agents can reference prior debates.

Can I use models other than OpenAI when self-hosting TradingAgents? Yes. Set the matching env var (ANTHROPIC_API_KEY, GOOGLE_API_KEY, DEEPSEEK_API_KEY, OPENROUTER_API_KEY, etc.) and pick the provider in the home-page form or pass llm_provider in the API body.

How long does a single analysis take to run on Railway? With one debate round and gpt-4o-mini quick-think, expect 60–180 seconds. Heavier models or higher debate counts can push this past 5 minutes — bump RAM and the uvicorn timeout if you hit edge timeouts.

Is this template safe to use for live trading decisions? The upstream project is research-grade and explicitly states it is not financial, investment, or trading advice. Treat the output as a reasoning artifact, not a signal.