Run scalable, production-ready AI assistants on Amazon Web Services for as little as $10/month.
Today we’re releasing something we’ve been working toward for a long time: AWS AI Agent Bus, a fully open source infrastructure framework for running AI agent meshes on Amazon Web Services. With this project, our goal is to make it practical—and affordable—for developers and businesses to deploy production-ready agent systems without months of trial-and-error infrastructure work.
Why We Built AWS AI Agent Bus for Scalable AI Assistants
AI assistants are evolving quickly. They’re no longer just toys that answer questions; they’re beginning to act like teammates. That means they need reliable systems underneath them. They need a way to remember things, to pass files back and forth, to coordinate through events, to handle multi-step workflows, and to do all of it without blowing up a small team’s budget.
Right now, the options are pretty limited. Either you stitch together some fragile glue code that doesn’t hold up under real workloads, or you buy into an enterprise-grade system that’s too expensive for startups and independent teams. We wanted a third option: something robust enough for production, simple enough to get running quickly, and cost-conscious enough to scale without stress.
What We Built
AWS AI Agent Bus has three major parts, designed to work together.
The first is a Model Context Protocol (MCP) server, which gives AI assistants standardized ways to talk to AWS. It supports DynamoDB for persistent key-value storage, S3 for managing files and artifacts, EventBridge for coordinating through events, and a structured timeline system for logging what happens when. It even gives you two options for interfaces—stdio or HTTP—depending on how you want your agents to connect.
The second piece is the Claude Agent Orchestration System. Think of this as a multi-agent team with defined roles. There’s a Conductor that sets the plan, Critics that check for safety and approval, and Specialists who bring deep knowledge of frameworks like React, Django, or Terraform. Behind them is a memory system that ties everything together, combining key-value storage, timeline history, and vector embeddings for contextual recall.
The third is a set of Terraform modules that define the infrastructure itself. We designed them as workspaces you can pick up depending on your needs. A small workspace spins up DynamoDB, S3, and EventBridge and costs about ten dollars a month. A medium workspace adds orchestration with ECS and Step Functions. And a large workspace brings in heavier options like Aurora with pgvector and analytics capabilities for teams pushing the limits.
Why It Matters
For developers, the benefit is speed. You don’t need to spend weeks wiring up infrastructure or worrying about hidden edge cases. You can deploy a production-ready stack in minutes, configure it entirely through environment variables, and rely on built-in error handling, monitoring hooks, and graceful shutdowns.
For businesses, the benefit is flexibility. You can start at just ten dollars a month and scale as your needs grow, without worrying about vendor lock-in or expensive licensing fees. Because it’s open source, there’s no barrier to entry—you can test, adapt, and extend the system however you like.
And for the AI community, this is a chance to work from shared patterns. Agent mesh architectures are still new territory. By releasing this project under the MIT license, with complete source code and professional documentation, we hope to give builders the building blocks they need to experiment safely and effectively.
Designed for Production from Day One
Cost optimization was a guiding principle from the very beginning. The small workspace, for example, comes in at around ten dollars a month: DynamoDB on-demand pricing typically falls between one and five dollars, S3 adds another one to three, and EventBridge runs about one to two. That’s it—you can run serious agent infrastructure for the price of a few cups of coffee.
We also obsessed over developer experience. Everything is driven by environment variables, so there’s no hardcoding. Error handling and logging are comprehensive. Health endpoints and monitoring hooks are baked in. And because we treat production readiness as non-negotiable, the system follows AWS security best practices throughout. You can have your first agent bus running in under ten minutes.
What You Can Do With It
We’ve already seen teams use AWS AI Agent Bus for a wide range of applications. Some are running AI-powered DevOps, where agents monitor deployments and handle incidents. Others are tackling content management, letting assistants process and distribute information at scale. We’ve also seen it applied to workflow automation, where agents coordinate complex multi-step processes, and to data pipelines, where AI helps manage ETL and analysis workloads.
Open Source and Moving Forward
We’re releasing AWS AI Agent Bus under the MIT license because we believe infrastructure like this should be open. You’ll find complete source code with no proprietary dependencies, full documentation, contribution guidelines, and examples you can run right away. It’s also CI/CD-ready, so you can slot it into professional development workflows without friction.
We’re already using the system in production, and we’re not stopping here. Upcoming work includes better monitoring and observability, more MCP integrations, additional Terraform workspaces, and ongoing improvements to performance and cost.
Get Involved
The repository is live at github.com/Baur-Software/aws-ai-agent-bus. Inside you’ll find the MCP server, Terraform infrastructure modules, and our .claude agent configs.
If you’re building AI-powered applications, experimenting with agent mesh architectures, or just curious about what production-ready AI infrastructure looks like on AWS, we’d love to see what you build. Reach out on LinkedIn or open an issue on GitHub—we’re looking forward to the conversation.