• 

Made by Darshon Singh

Inspiration

I wanted to make a tool that would help me in the startup process, so I thought what my next steps were. I wanted to be able to purely focus on the product development side, while leaving all of the business side to AI.

What it does

MBAi transforms Claude into your entire business operations team through a visual, no-code interface. Instead of juggling multiple tools or hiring expensive specialists, users can build AI-powered departments in seconds by dragging and dropping agents onto a canvas.

Each agent is powered by Claude and can be customized with specific instructions and capabilities. The CEO Assistant acts as your command center, answering questions about company progress, tracking completed work, and providing insights by querying the built-in knowledge base using semantic search.

When you execute a workflow, agents process tasks in real-time with live streaming feedback. Every completed task automatically updates the company knowledge base, creating an evolving memory system that gets smarter over time. Agents have access to built-in tools like web search and industry research, plus the ability to integrate custom webhooks for specialized operations.

Whether you're a solopreneur managing customer service, a startup founder coordinating product launches, or a small business handling multi-step processes, MBAi gives you the power of an AI workforce that remembers context, learns from past work, and scales instantly.

Business Value

The Problem: Small businesses and solopreneurs face a critical bottleneck: they need specialized expertise across multiple domains—marketing, customer service, data analysis, content creation, operations—but can't afford to hire full-time specialists for each function. According to the U.S. Bureau of Labor Statistics, the median salary for a business operations specialist is $78,000/year. For a small business to build a basic team of 5 specialists across different functions, that's $390,000 in annual salary costs alone, not including benefits, office space, equipment, or training.

The alternative—using multiple SaaS tools—creates its own problems: fragmented workflows, data silos, integration headaches, and a patchwork of subscriptions that still require manual coordination. The average small business uses 40+ SaaS applications, spending $5,400-$20,000 per year on software, and employees waste 2-3 hours daily switching between tools and searching for information.

The Solution: MBAi collapses this entire operational overhead into a single, unified AI workforce. For the cost of Claude API calls (typically $50-200/month for a small business), you get:

• Instant Expertise Across Domains: Create specialized agents for marketing, customer service, research, data analysis, content creation, and more. Each agent brings Claude's broad knowledge base to their specific role.

• Zero Hiring Friction: Spin up a new "department" in 30 seconds. Need a competitive analysis team? Drag, drop, configure, done. Compare that to a 3-month hiring process for a human analyst.

• Perfect Memory: Unlike human employees who forget context or previous work, the knowledge base ensures every agent has access to all past decisions, completed tasks, and learned insights. No more asking "didn't we already research this?"

• 24/7 Availability: Your AI workforce doesn't sleep, take vacations, or call in sick. Execute workflows at 3 AM or handle customer inquiries on weekends without paying overtime.

• Infinite Scalability: Need to process 100 customer inquiries instead of 10? Just execute the workflow 100 times. No new hires, no training period, no scaling pains.

Real-World Impact:

Solopreneur Consultant: Before MBAi, a solo consultant might spend 15 hours/week on non-billable work: researching industries, writing proposals, analyzing data, managing social media. With MBAi, they create a Research Agent, a Writing Agent, and an Analytics Agent. Those 15 hours drop to 3 hours of reviewing AI output. That's 12 hours back—enough to take on 2 more clients, increasing annual revenue by $50,000+ while working less.

Small E-commerce Business: A 3-person e-commerce company handles customer service, product research, inventory decisions, and marketing content. Each person juggles multiple roles, leading to burnout and inconsistent quality. With MBAi, they build a Customer Service Agent (handles common inquiries), a Product Research Agent (analyzes trends and competitors), and a Content Agent (writes product descriptions). The team goes from reactive firefighting to strategic oversight, focusing on high-value decisions while AI handles the execution layer. Customer response time drops from 4 hours to 15 minutes. Product research that took 6 hours now takes 30 minutes of review time.

Startup Founder: A technical founder building a B2B SaaS product needs to conduct market research, create content, analyze user feedback, and coordinate product launches—but they want to focus on engineering. MBAi becomes their entire "business operations team." They create workflows for each business function and execute them as needed. Instead of spending 20 hours/week on business tasks, they spend 5 hours reviewing AI output. The product ships 3 weeks earlier because the founder isn't context-switching constantly.

Competitive Advantages:

1. Visual Workflow Design: Unlike chat-based AI tools (ChatGPT, Claude.ai) that lose context between sessions, MBAi provides a persistent, visual workspace where workflows are reusable assets. Unlike no-code automation tools (Zapier, Make), MBAi brings intelligent reasoning to every step, not just mechanical API calls.

2. Built-In Memory: The knowledge base differentiates MBAi from ephemeral AI assistants. Every completed task enriches future executions. Competitors either have no memory or require complex external integrations.

3. True Multi-Agent Orchestration: Tools like AutoGPT or AgentGPT focus on single-agent autonomy. MBAi enables coordinated multi-agent workflows with specialized roles, mirroring how real businesses organize expertise.

4. No-Code + Pro-Code: Non-technical users get a point-and-click interface. Technical users can integrate custom tools via webhooks. This hybrid approach captures both market segments.

Market Opportunity:

The global small business market includes 33.2 million businesses in the U.S. alone, with 27 million being solopreneurs or having fewer than 5 employees. If just 0.1% adopt MBAi at $50/month average revenue, that's a $19.9 million annual market. The broader SMB automation market is projected to reach $13.7 billion by 2028.

But the real opportunity isn't competing with existing SaaS tools—it's unlocking new capabilities for businesses that previously couldn't afford sophisticated automation. MBAi democratizes access to AI-powered operations, similar to how Shopify democratized e-commerce or Canva democratized design.

ROI Example:

A small consulting firm with 5 employees currently spends:

• $80,000/year on a junior operations coordinator
• $10,000/year on SaaS tools (CRM, project management, research databases)
• 30 hours/week of senior staff time on coordination tasks ($150/hour × 30 hours × 52 weeks = $234,000 in opportunity cost)

With MBAi:

• $2,400/year on Claude API costs
• 5 hours/week of senior staff time reviewing AI output ($150/hour × 5 hours × 52 weeks = $39,000)
• Total savings: $282,600/year

Even at a conservative 25% of that theoretical savings ($70,650), the ROI is undeniable.

Why Now:

Three converging trends make this the perfect time for MBAi:

1. AI Capability Threshold: Claude 3.5 Sonnet is the first model that's reliable enough for business-critical tasks while being affordable enough for small businesses. Earlier models were too expensive or too unreliable.

2. Remote Work Normalization: Post-pandemic, businesses are comfortable with "team members" who aren't physically present. The psychological barrier to AI coworkers is lower than ever.

3. No-Code Expectations: A generation of business users now expects visual, intuitive tools. If Figma can make design accessible and Webflow can make web development accessible, AI orchestration should be accessible too.

MBAi isn't just a hackathon project—it's a blueprint for how small businesses will operate in the AI-first economy. Every business will need AI operations. MBAi makes it possible without requiring data scientists, ML engineers, or massive budgets.

How we built it

CLAUDE CODE!!

Frontend Architecture:

• Built with Next.js 16 using the App Router for modern React patterns
• React Flow powers the visual workflow canvas with drag-and-drop agent nodes
• Zustand manages global state for agents, executions, and workflows
• TypeScript ensures type safety across the entire application
• Tailwind CSS with custom color palette for consistent styling

AI & Backend:

• Dual-model strategy: Claude Haiku powers department agents for fast, cost-effective execution, while Claude Sonnet 4.5 runs the CEO assistant for complex reasoning
• Custom RAG (Retrieval-Augmented Generation) system built from scratch using TF-IDF vector embeddings
• Document chunking based on semantic boundaries (H1, H2, H3 headers) with metadata tagging
• Cosine similarity search with dynamic thresholds for intelligent context retrieval
• 10-turn tool calling loop with retry logic and error handling

Real-Time Features:

• Server-Sent Events (SSE) with ReadableStream encoders for live execution streaming
• Atomic node status updates on the canvas as agents complete tasks
• Immediate vector cache invalidation for instant knowledge base search availability

Data & Persistence:

• Knowledge base stored in company.md with automatic updates after task completion
• LocalStorage for workflow state and execution history
• Edge-first deployment on Vercel for global low-latency access

Tool Integration:

• Built-in tools: web search (Brave API) and industry research (Wikipedia API)
• Custom tool system following Model Context Protocol patterns
• JSON schema validation for tool definitions
• Sandboxed execution with proper authentication (Bearer, API Key)

Challenges we ran into

Claude code running out of usage.

Building RAG from Scratch: The biggest challenge was implementing a custom RAG system without relying on heavy vector database libraries. We had to understand TF-IDF weighting, document tokenization, stop word filtering, and cosine similarity calculations. Getting the chunking strategy right took multiple iterations—we needed chunks large enough to be meaningful but small enough to be relevant. The semantic boundary detection using headers was key to making search actually useful.

Real-Time Streaming Architecture: Making Server-Sent Events work reliably with Next.js API routes required careful stream management. We had to handle cases where agents called multiple tools in sequence, ensure proper cleanup on connection drops, and make sure the UI stayed responsive while receiving rapid updates. The ReadableStream encoder pattern was tricky to get right.

Tool Calling Orchestration: Claude's tool calling API is powerful but complex. We built a recursive loop that handles tool results, formats them properly for the next turn, detects when agents are done, and gracefully handles errors. Getting the retry logic right—especially for external API failures—took significant debugging.

Knowledge Base Consistency: Keeping the vector store in sync with the markdown file while supporting real-time updates was challenging. We needed to invalidate caches at the right time, prevent duplicate chunks, and ensure the CEO assistant always had access to the latest completed work without fetching stale data.

State Management Across Executions: Maintaining context across multiple workflow runs while keeping the UI snappy required careful state design. We used Zustand for its simplicity but had to structure stores to avoid unnecessary re-renders when streaming updates came in.

Accomplishments that we're proud of

Complete Multi-Agent System in 24 Hours: We built a fully functional multi-agent orchestration platform from scratch during the hackathon. Users can create agents, design workflows visually, execute them with real-time feedback, and query results through an intelligent assistant. Everything works end-to-end.

Custom RAG That Actually Works: Our TF-IDF implementation isn't just a proof of concept—it delivers genuinely relevant search results. The CEO assistant can answer complex questions like "What did the Marketing team work on last week?" and get accurate, context-aware responses by querying the vector store.

Seamless Tool Integration: The tool calling system handles both built-in tools and custom webhooks elegantly. Agents can search the web, fetch industry data, and call external APIs with authentication, all orchestrated automatically by Claude with proper error handling and retries.

Beautiful, Intuitive UX: Despite being a hackathon project, the UI is polished and professional. The React Flow canvas feels like a native design tool, the CEO chat interface is clean and responsive, and the real-time execution panel provides clear feedback on what's happening.

Production-Ready Deployment: The application is deployed on Vercel with edge functions, properly handles state persistence, and has been tested with real workflows. It's not just a demo—it's actually usable.

What we learned

Learned a TON. First time using Git, Vercel, and Claude code. Claude Code is immensely powerful, although the limits aren't great. Exciting that making an app this complex can be made in under 24 hours.

Claude API Patterns: We learned how to maximize Claude's tool calling capabilities, structure prompts for multi-turn conversations, handle streaming responses efficiently, and use the Haiku vs Sonnet models strategically based on task complexity and cost.

Multi-Agent Orchestration: Building a system where multiple AI agents work together taught us about context management, result aggregation, and designing clear agent roles. We learned when to parallelize vs sequence agent execution.

Real-Time Web Applications: Implementing SSE for live updates taught us about long-lived HTTP connections, stream cleanup, browser compatibility, and building responsive UIs that handle rapid state changes without janky re-renders.

Product Design for AI Tools: We learned how to make AI capabilities accessible to non-technical users through visual interfaces, how to provide feedback for inherently uncertain AI operations, and how to design workflows that feel deterministic even when powered by LLMs.

What's next for MBAi

Workflow Templates: Pre-built workflow templates for common business scenarios—customer service routing, market research reports, content creation pipelines, data analysis workflows. Users can start with a template and customize rather than building from scratch.

Agent-to-Agent Communication: Enable agents to communicate and collaborate on complex multi-step tasks. An agent could delegate subtasks to specialists, request information from other departments, and coordinate work without human intervention.

Tool Marketplace: A community marketplace where users can share and sell custom tools, workflow templates, and agent configurations. Imagine a library of pre-integrated APIs for CRM, marketing automation, analytics platforms, and industry-specific tools.

Advanced Visualization: Show agent reasoning processes, tool call sequences, and decision trees visually. Let users "peek inside" to understand why an agent made certain choices and debug workflows more effectively.

Workflow Version Control: Git-like versioning for workflows with branches, commits, and rollback. Teams could collaborate on workflow design with proper change tracking and approval processes.

Agent Training from Feedback: Use the pending actions system to collect feedback and fine-tune agent behavior. Approved actions become positive examples, rejected ones become negative examples, creating a continuous improvement loop.

Enterprise Features: Team collaboration with role-based permissions, audit logs for compliance, SSO integration, and private knowledge bases. Make MBAi enterprise-ready for larger organizations.

Multi-Modal Agents: Support agents that can process images, PDFs, and audio. Imagine a document processing department that extracts data from invoices or a quality assurance team that analyzes screenshots.

Cost & Performance Analytics: Detailed dashboards showing token usage, execution times, success rates, and cost per workflow. Help users optimize their AI operations with data-driven insights.

API & SDK: Programmatic access to MBAi workflows so developers can trigger them from their own applications, integrate with existing systems, and build custom interfaces on top of the platform.

Built With

• claude
• next.js
• typescript