Running TradingAgents locally

Tauric TradingAgents is an open source project that uses several AI agents to work through investment research. The agents take on different roles: analysts, researchers, a trader, risk reviewers, and a portfolio manager. The project’s README explains how they work together. This post covers how I connected the framework to cloud and local models, with a rate limiter to keep requests under control.

The usual setup needs API keys for cloud models. I already had AI subscriptions and wanted to use whatever allowance I had left in those instead of paying separately for API calls. So I put a local model gateway on my MacBook. It lets me inspect the traffic and limit requests, and I can switch providers without changing the TradingAgents code.

Neither this article nor the TradingAgents project is financial advice. I treat the output as AI-generated research notes that need checking, rather than a reason to buy or sell.

The Architecture of my deployment

My local setup looks like this:

CLIProxyAPI gives TradingAgents an OpenAI-compatible endpoint. Behind that endpoint, I can use Gemini or Codex (or any other OpenAI-compatible LLM provider) through my subscriptions, or connect a local model. TradingAgents handles the research; CLIProxyAPI handles access to whichever provider I choose.

The rate limiter sits between TradingAgents and the gateway. Several agents can send requests at once, which can trigger provider limits, especially when using a subscription. The small token-bucket proxy lets me control how quickly those requests reach the gateway.

Before you start

I used a MacBook with Homebrew and Python 3.10+. You can run the setup on another machine, though the Homebrew and macOS service commands below will need adapting.

I kept the Python environment in a .venv folder inside the project so it was easy to find and inspect. If you prefer Conda or Docker, the upstream README covers those options.

1. Install TradingAgents in a local virtual environment

Clone the Tauric TradingAgents repository, then create a virtual environment and install the project:

git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgents
python3 -m venv .venv
.venv/bin/python -m pip install --upgrade pip
.venv/bin/python -m pip install .

To run tradingagents from any shell, add a symlink in a directory on your PATH. For example, with ~/.local/bin:

ln -s "$PWD/.venv/bin/tradingagents" ~/.local/bin/tradingagents

Make sure ~/.local/bin is on your PATH for this shortcut to work.

2. Install CLIProxyAPI and sign in to your providers

CLIProxyAPI is the part that connects to an existing subscription, such as Google AI Pro or OpenAI Codex. TradingAgents sends its requests to the local gateway.

Install CLIProxyAPI with Homebrew:

brew install cliproxyapi

Sign in to each provider you want to use. For Gemini, this opens a browser login:

cliproxyapi -login

For Codex, use the device login flow:

cliproxyapi -codex-device-login

Then restart the service:

brew services restart cliproxyapi

CLIProxyAPI saves the authentication tokens as .json files in ~/.cli-proxy-api.

With my Homebrew installation, the configuration file is here:

/opt/homebrew/etc/cliproxyapi.conf

I’m also sharing an example file in my config repo: https://github.com/atsamour/tradingagents-proxy-config

The template binds CLIProxyAPI to localhost and enables its Management Center. It also defines the client key that TradingAgents will send with requests; you’ll put that key in the TradingAgents .env file later. Optional providers that use API keys stay disabled until you supply their keys.

There are two local keys to keep track of:

• a client API key for requests from TradingAgents to CLIProxyAPI
• a management key for CLIProxyAPI’s Management Center

Generate both keys and save their values. You’ll need them again when configuring TradingAgents and signing in to the Management Center:

CLIENT_KEY="tradingagents-$(openssl rand -hex 24)"
MGMT_KEY="mgmt-$(openssl rand -hex 24)"
echo $CLIENT_KEY
echo $MGMT_KEY

Back up the current configuration, replace the template’s placeholders with your keys, then install it and restart CLIProxyAPI:

cp /opt/homebrew/etc/cliproxyapi.conf /opt/homebrew/etc/cliproxyapi.conf.backup
cp config/cliproxyapi.conf.example /tmp/cliproxyapi.conf
perl -0pi -e "s/tradingagents-REPLACE_WITH_GENERATED_CLIENT_KEY/$CLIENT_KEY/g; s/mgmt-REPLACE_WITH_GENERATED_MANAGEMENT_KEY/$MGMT_KEY/g" /tmp/cliproxyapi.conf
install -m 600 /tmp/cliproxyapi.conf /opt/homebrew/etc/cliproxyapi.conf
brew services restart cliproxyapi

If your Homebrew directory requires elevated permissions, use them only for the install step.

Open the Management Center at:

http://127.0.0.1:8317/management.html#/login

Sign in with the MGMT_KEY you generated above.

I keep CLIProxyAPI and its Management Center on the same machine as TradingAgents. You could also run the gateway on a separate machine and configure network access to it.

3. Point TradingAgents at the local gateway

TradingAgents lets you override its model settings with environment variables. Create a .env file in the TradingAgents repository:

# .env (Local Overrides)
OPENAI_API_KEY=CLIENT_KEY
TRADINGAGENTS_LLM_PROVIDER=gemini
TRADINGAGENTS_LLM_BACKEND_URL=http://127.0.0.1:8318/v1
TRADINGAGENTS_DEEP_THINK_LLM=gemini-3.1-pro-preview
TRADINGAGENTS_QUICK_THINK_LLM=gemini-3.1-flash-lite-preview
ALPHA_VANTAGE_API_KEY=your-alphavantage-key-here
CLIPROXYAPI_MANAGEMENT_KEY=your-generated-$MGMT_KEY

Get the ALPHA_VANTAGE_API_KEY here: https://www.alphavantage.co/support/#api-key. It is a stock market data retrieval API.

Use the CLIENT_KEY value you generated in the previous step. The backend URL points to the rate limiter on port 8318, a small Python script built with aiohttp. Sending requests through it lets me control how quickly the agents use my subscription.

To find model names for TRADINGAGENTS_DEEP_THINK_LLM and TRADINGAGENTS_QUICK_THINK_LLM, run:

scripts/app/tradingagents_models.sh list

Or query the model endpoint directly:

curl -fsS \
  -H "Authorization: Bearer mgm-token-here" \
  http://127.0.0.1:8318/v1/models

Start the local rate limiter with:

scripts/app/cliproxyapi_rate_limiter_service.sh submit

4. Set up the rate limiter

CLIProxyAPI already handles retries, provider routing, and cooldowns. I also wanted a requests-per-minute limit that I could set myself, right at the endpoint TradingAgents calls.

The limiter is a small aiohttp reverse proxy in:

scripts/app/cliproxyapi_rate_limiter.py

It listens on 127.0.0.1:8318 and forwards requests to 127.0.0.1:8317. You can check its health at:

http://127.0.0.1:8318/healthz

I use these conservative defaults:

10 requests per minute, burst 3

To change them, edit:

config/cliproxyapi-rate-limiter.env

To keep the limiter running as a macOS service through launchd, use:

scripts/app/cliproxyapi_rate_limiter_service.sh submit
scripts/app/cliproxyapi_rate_limiter_service.sh status

For a foreground session that you start and stop yourself, use:

scripts/app/cliproxyapi_rate_limiter_service.sh start
scripts/app/cliproxyapi_rate_limiter_service.sh status
scripts/app/cliproxyapi_rate_limiter_service.sh stop

When a request includes a bearer token, the limiter uses its API key to choose a bucket. That keeps request limits separate if several local clients use the gateway with different keys.

5. Switch models without editing .env by hand

It’s easy to put a model name in .env that TradingAgents accepts but the gateway can’t serve. The helper script checks CLIProxyAPI’s current model list before updating the file, so you can catch that mistake before starting a run.

List available models:

scripts/app/tradingagents_models.sh list

Check which models TradingAgents is currently set to use:

scripts/app/tradingagents_models.sh show

Use the Gemini preset:

scripts/app/tradingagents_models.sh preset gemini

Use the Codex preset:

scripts/app/tradingagents_models.sh preset codex

Or choose the model IDs yourself:

scripts/app/tradingagents_models.sh set \
  --deep gemini-3.1-pro-preview \
  --quick gemini-3.1-flash-lite-preview

TradingAgents may warn that your Gemini or Codex model isn’t in its built-in OpenAI model catalog. That’s expected when using the proxy. The helper checks the gateway’s /v1/models response to confirm which models it can serve.

Here are a few watchlist runs using the Gemini and Codex presets:

scripts/app/run_stock_watchlist.sh --mode news --model gemini PLTR GOOGL RHM.DE
scripts/app/run_stock_watchlist.sh --mode potential --model codex PLTR GOOGL RHM.DE
scripts/app/run_stock_watchlist.sh --mode full --model gemini PLTR

To preview the command without starting the research run, add --dry-run:

scripts/app/run_stock_watchlist.sh --mode news --model gemini --dry-run PLTR

The wrapper saves reports under:

runs/tradingagents/<TICKER>/<DATE>/summary.md
runs/tradingagents/<TICKER>/<DATE>/reports/

Check the setup

Before starting a full research run, I use the doctor script to check the setup. It checks the Homebrew service and listening ports, then the Management Center, provider status, and model selection. It also checks that the CLI starts and sends a small completion request through the limiter:

scripts/setup/tradingagents_doctor.sh

This script checks if the ports are open, if the providers are authenticated, and runs a “smoke test” completion.

Run a research session

Once those checks pass, I start the TradingAgents CLI for an interactive session:

tradingagents

You can choose which analyst teams to run, for example market,news,fundamentals. Tickers can come from any market Yahoo Finance covers, including symbols such as AAPL, 0700.HK, or BTC-USD.

When I want to repeat the same research across a watchlist, I use the batch wrapper. It saves the results as Markdown:

scripts/app/run_stock_watchlist.sh --mode full --model gemini PLTR GOOGL

The reports go into runs/tradingagents/. I can read through the bull and bear arguments later without running the models again.

Final Thoughts

I started this because I wanted to run TradingAgents locally without paying for extra API calls. Keeping provider access in CLIProxyAPI and request limits in a separate proxy made the setup easier to manage.

If you’re trying something similar, I’d recommend putting a gateway between your agents and the providers. Being able to see the requests and slow them down when needed ensures that your AI subscription won’t get banned. It also gives you a place to start looking when an agent appears to be stuck or even review how agents operate based on the input they get.

As a next step I would like to run some benchmarks with past starting date (TradingAgenst support it!) and see if predictions are accurate and when not, analyse why. Will post in due time.

Please feel free to share your experience with useing TradingAgents and don’t hesitate to comment if yuou have any issues running them locally, as described here.

NOTES:

• Τhis was tested on TradingAgents v0.2.5
• All configs and scripts can be found here: https://github.com/atsamour/tradingagents-proxy-config