Building a Visual Workflow Ecosystem on MCP

by

TL;DR

  • Real-time pub/sub dashboard and UI
  • AWS Bedrock + Ollama AI support
  • Rust MCP server with multi-context isolation
  • Composite hub integrating any MCP servers you choose
  • 900 + MCP servers, 1,000 + workflow nodes
  • Event-driven via EventBridge
  • 11 importable, full-stack examples
  • Open-source, MCP-native alternative to Make/Zapier

Building a Visual, Executable Workflow Ecosystem

Most of my day used to bounce between whiteboards for design and automation tools for execution.
So I asked: why can’t they be the same thing?

Now, the whiteboard is the system. Each node on the canvas isn’t a diagram—it’s a live, executable component.

Example of the UI we’re working on. We have some work to do on connection still in the UI. This is the result of a query to our trained Claude Sonnet layer on the MCP workflow system.

The AWS AI Agent Bus: Multi-Context, Composite MCP

Our implementation—aws-ai-agent-bus/mcp-rust—isn’t a single-tenant adapter; it’s a multi-context, composite hub.

It can securely route requests across users, organizations, and environments while aggregating many MCP servers under one logical interface.

Composite Interface Highlights

  • Plug in any number of MCP servers (local, Docker, or public)
  • Chain tools across them in one workflow
  • Add / remove integrations without redeploying

Run Claude through Anthropic’s MCP. Hit AWS through the dasbhoard-server’s event bus. Call GitHub automation and trigger local utilities. All of this can be done from the same workflow.

That’s the power of a composite, context-aware MCP server.

Why Not Just Use Make or Zapier?

They’re convenient but closed: proprietary connectors, polling loops, and limited extensibility.
We’re open-source and event-driven—built on MCP for semantics, EventBridge for speed, and open code for trust. We want people to contribute to it because there isn’t a really good open-source solution in this space.

The AI-Native Advantage

Because tools follow MCP schemas, AI can execute instead of merely generating click-flows.

Ask Claude:

“Pull last month’s GA data, store in DynamoDB, and chart the trend.”

Claude orchestrates building out workflow nodes directly into the canvas.

Our conductor / critic / specialist agent pattern can optimize, debug, and regenerate workflows autonomously.

Real-Time Execution

EventBridge helps orchestrate the dashboard-server and updating the UI:

  • GitHub push → build starts instantly
  • Payment received → invoice sent
  • Sensor trigger → light adjusts

The dashboard updates live via WebSocket, giving you true real-time visibility.

Multi-Tenant Isolation & Ownership

Your workflows run inside your AWS account using S3, DynamoDB, EventBridge, and Secrets Manager.
Strict folder-level isolation keeps agent data separate per org/user:

S3: agents/
├─ organizations/{orgId}/
│  ├─ shared/
│  ├─ generated/{mcpId}/
│  └─ users/{userId}/
└─ users/{userId}/

Compliance, sovereignty, and predictable cost—all under your control.

Predictable Cost Model

No per-step billing.
You pay for AWS usage only:

TierTypical CostComponents
Small≈ $10 / moDynamoDB + S3 + EventBridge + Bring your own LLM
Medium$50–100 / mo+ ECS + Step Functions + Bedrock
Large$200–500 / mo+ Aurora + analytics stack

Architecture: Git-Inspired Ops

  • Continuous autosave + atomic rollback
  • Fork / import / export workflows like repos
  • RBAC contexts (personal ↔ org) on DynamoDB & S3
  • Infrastructure defined as Terraform (/infra)

Why Rust

JavaScript’s flexibility turned risky at scale. Rust gives compile-time guarantees, deterministic memory free, fearless concurrency, and type-level AWS quotas.

Performance Wins

  • Startup 800 ms → 50 ms
  • Latency 30 ms → 5 ms
  • Memory 150 MB → 8 MB
  • Concurrency 100 → 1000+ requests

Correctness > speed of typing. Once it compiles, it just works. You can drop it into Claude Code, set up .env file, and type /mcp aws. You’re connected and sharing context with your team now.

MCP Core Tools

  • KV Memory: kv_get, kv_set (< 5 ms)
  • Artifacts: S3-backed get/put/list
  • Events: send, query, analytics, alerts, health check
  • Discovery: 900 + MCP servers available in the registry

Every action emits an event → complete audit trail, analytics, and composability.

UI Updates

We treat workflow nodes like installable NPM packages. This means that you can upload custom workflow nodes into the node registry in the dashboard server. Take a look at the current nodes we have now for examples.

Example Workflows Included

The repository ships with 11 importable workflows that demo the architecture. Full-stack automations built on the composite MCP hub.

🧩 Infrastructure & DevOps

terraform-infra.yaml — Provisions AWS infrastructure through the canvas using Bedrock, EventBridge, and Secrets Manager.

💰 Finance & BI

quickbooks-crm-dashboard.yaml — Correlates invoices ↔ leads for live ROI dashboards, publishes to S3 and Slack.

📈 Analytics & Content

google-analytics-content-gen.yaml — Ingests GA data, stores in DynamoDB, and invokes AI for keyword-driven blog generation.

🏠 Smart Home

home-assistant-automation.yaml — Syncs Home Assistant triggers to AWS Lambda for lights, alerts, and video events.

🤖 AI Orchestration

claude-code-chain.yaml — Uses Claude Code + Bedrock to auto-generate workflows from text prompts.

All examples live in /workflow-examples/ and can be imported, forked, and extended instantly.

Quick Start

# Dev mode
npm run dev
# or provision infra
WS=small ENV=dev npm run tf:init && npm run tf:apply

Open the dashboard at http://localhost:5173, connect a few MCP servers, and start designing.

Links & Community

Contributions

I welcome contributions to this project. It is early but we feel the code is stable enough that most of what is needed is improving the canvas and finalizing implementation of what is mocked in there. This project started humbling out of a need for collaboration with engineers leveraging LLM’s in our teams. Who wanted to see how far we can push Claude Code MCP to be a great collaboration tool. A ‘pivot’ to workflows was a natural evolution for us as a next step. When we were discussing the use of this internally and our advisory staff said, “I want that too!”.

Workflows and systems are evolving fast. What we need more of now than ever is to capture context in them. It’s very easy to go to Make or n8n. Anyone can create something to automate a business process. However, it’s easy to forget why such an automation was made the way it was. APIs change without much warning and they break. Authentication times out. They perform unknowingly bad transforms, etc. That’s why we thought about things like Sticky Notes for the canvas. These sticky notes accept html, data inputs, and emit data outputs so you can be very clever with them too.

Leave a Reply