AI Agent with MCP Server on AWS Lambda

This repo contains a sample implementation of user-aware AI Agent and MCP Server running on AWS Lambda. The sample implements an AI-based Travel Agent for a fictitious corporation AcmeCorp.

See two below artifacts to better understand the expected Agent behavior.

• ./lambdas/travel-agent/agent_config.py for Agent's system prompt.
• ./lambdas/bookings-mcp/tool-get-travel-policies.js for corporate travel policies the Agent is bound by.

(Skip to Running the project)

Frameworks and services

• AI Agent is implemented using Strands Agents SDK
• MCP Server is implemented using the official MCP TypeScript SDK and Lambda Web Adapter
• Amazon API Gateway is used to front AI Agent and MCP (see architecture diagram).
• Amazon S3 is used to persist AI Agent session state.
• Amazon Bedrock and Claude Haiku 3.5 are used for inferrence.
• Gradio is used for Web UI.

Authentication and Authorization

• User authentication is handled by Amazon Cognito
• During terraform apply/cdk deploy two users will be created in the Cognito User Pool - Alice and Bob. Use these users to login, as described below.
• Both AI Agent and Bookings MCP domains require JWT (bearer token) authorization, as enforced by API Gateway Lambda authorizers.
• AI Agent expects JWT issued by Cognito. The subject of this JWT is the user. Validation is done in the Agent Authorizer using JWKs generated by Cognito during user authentication.
• Bookings MCP expects JWT generated by Strands Agent. The subject of this JWT is the agent with user_id being an extra claim. This token is signed by a shared secret. Validation is done in the MCP Authorizer using shared secret.
• To state explicitly - at no point in time user's identity is inferred from LLM's response. It is ALWAYS propagated using JWTs.

User-awareness and session management

• Both AI Agent and MCP Server are fully stateless, they do not maintain any internal state.
• Both AI Agent and MCP Server are aware of the current user and process requests within the user authorization context.
• AI Agent externalizes its state to S3, using the S3SessionManager class provided by Strands SDK.
• New Strands Agent instance is created on every request with state that belongs to the current user.
• MCP Clients attached to the AI Agent are created per-user, holding user-specific JWT to communicate with MCP Server.
• MCP Server propagates authorized user context into all tools, making MCP tools user-aware.

Tools

Implemented in agent

• [Agent] get-user-location resolves user's IP address to users location.
• [Agent] get-todays-date tool returns today's date.

Implemented in MCP Server

• [MCP] get-travel-policies returns AcmeCorp corporate travel policies.
• [MCP] get-available-cars returns a list and categories of cars available to rent.
• [MCP] book-car books a car rental.
• [MCP] book-hotel books a hotel.

Running the project

The project is using arm64 architecture by default for better cost efficiency. Update IaC files if you want to change it to x86.

Prereqs

• AWS CLI, Git, Docker
• Terraform or CDK
• Node.js installed
• Access to us.anthropic.claude-3-5-haiku-20241022-v1:0 in us-east-1. Or update Agent configuration to use a different model/region.

Clone the project

git clone [email protected]:sjpandruju/aws-strands-agent.git

Install dependencies

(cd lambdas/agent-authorizer && npm install)
(cd lambdas/mcp-authorizer && npm install)
(cd lambdas/bookings-mcp && npm install)

Deploy to AWS with CDK or Terraform

Terraform:

cd terraform 
terraform init
terraform plan
terraform apply

CDK:

npm install
cdk deploy

Make sure to carefully review required IAM permissions.

Run post-deploy script

Terraform:

# Make sure you're in the {project_root}/terraform directory
./prep-web.sh
cd ..

# Make sure you're in the {project_root}/ directory
./prep-web.sh

Running this script will

1. Update Cognito users Alice and Bob with passwords. The password is Passw0rd@. You will need it to login.
2. Update {project_root}/web/.env file with configuration required for running Web UI.

Run Web UI

cd web
python3 -m venv .venv             # Create virtual environment
source .venv/bin/activate         # Active virtual environment
pip install -r requirements.txt   # Install dependencies
python app.py                     # Start Web UI app

Open http://localhost:8000/chat/ in your browser

Login

You will be redirected to Cognito-hosted login screen.

• Username: Alice or Bob
• Password: Passw0rd@

Once logged in, start asking your AI Agent questions about corporate travel booking, for example

• What cars can I rent?
• Book me a trip to Paris
• I need to go to Dallas for three days starting next Monday, I also need a car
• Can I rent a Mercedes?

Clean-up

Terraform:

terraform destroy

CDK:

cdk destroy

License

This library is licensed under the MIT-0 License. See the LICENSE file.