Claude Stonks Agent

An example langchain+langgraph agent using Anthropic's Claude 2.1 on AWS Bedrock to demonstrate function calling.

Creates answers using langchain tools that expose the alphavantage.co stock market API.

Also has a very basic Streamlit UI.

Setup

Install

This is just using a python virtual environment.

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Environment Variables

You'll need an AWS account with Claude 2.1 enabled on Bedrock. Setup a user to access this and set the correct environment variables for boto3 sessions to use it by default. For example create a profile in ~/.aws/config and set the AWS_PROFILE environment variable.

You'll also need a alphavantage.co API key which is available here: get free API key. The free key is limited to 25 requests a day, so to do more experiementation you'll need a premium key: buy premium key.

We use a .env file and python-dotenv to load it. Create a .env file with something like:

AWS_PROFILE=your-profile-name
ALPHA_VANTAGE_API_KEY=your-key

Running the UI

You should just be able to run the Streamlit UI with:

streamlit run ui.py

Notable points

• Claude 2.1 tool/function calling is mentioned as being in "early access" so almost certainly will change. (docs)
• The prompts for tool usage used here are based on anthropic-tools that also makes clear: This SDK is Currently in Alpha. We promise no ongoing support. It is not intended for production use.
• I found the Bedrock client from langchain needed a little customization to work well. See the ClaudeBedrock implementation in claude.py. Do let me know if there are better ways of doing this.
  • The fix_prompt function to change the AI: to Assistant: prefixes.
  • Setting the </function_calls> stop token in the model kwargs rather than then usual langchain way of binding it as a stop token on the model.
• Isn't this stonks thing a bit silly? Yep.

Missing Bits

• Based on the anthropic-tools example it looks like multiple function calls are possible in a single response, but this is not yet explored.
• There is no tool error handling. Again, there is some examples of this in anthropic-tools.