Hive

Hive connects AI to all your apps in one unified chat. Describe what you need and it intelligently routes requests to 300+ services, chaining multiple actions together automatically.

Comment

Inspiration

I've always been frustrated by the fragmentation of our digital tools. We have Asana for tasks, Gmail for email, Slack for messaging, Notion for docs—and we're constantly switching between them. The promise of AI agents is that they can work for us, but most agent implementations are siloed: one agent, one tool, one conversation.

I wanted to build something different: a unified interface where AI can intelligently access all your connected services at once—not just one tool at a time. The vision is an agent ecosystem that genuinely works for you, handling the coordination so you don't have to.

When I discovered the Composio Tool Router and MCP (Model Context Protocol), I knew these were the foundations I needed.

What it does

Hive is a unified chat interface where you talk to AI that has access to all your connected services. Instead of switching between apps, you simply describe what you need:

  • "Create a task in Asana for Q1 budget review"
  • "Check my calendar and send John an email with available times"
  • "Pull the latest GitHub issues and add them to my Linear board"

The AI intelligently routes your requests to the appropriate services, executes the actions, and returns results—all in a familiar chat experience. You can even enable code execution skills for data analysis, document generation, and more.

Key features:

  • Multi-model support: Switch between OpenAI, Anthropic, Google Gemini, Mistral, Perplexity, xAI, and local Ollama models
  • 300+ app integrations: Connect Gmail, Asana, Slack, Notion, GitHub, Linear, Google Calendar, and more via Composio
  • Code execution skills: Run Python in sandboxed E2B environments for data analysis, chart generation, and document creation
  • Projects: Organize chats and configure which skills/integrations are active per project
  • Full observability: LLM tracing with Arize Phoenix and OpenTelemetry

How I built it

Frontend:

  • Next.js 16 with React 19 and Turbopack
  • Tailwind CSS 4 with Radix UI/Shadcn components
  • Zustand for state management
  • Motion for animations

AI Orchestration:

  • Vercel AI SDK v6 with multi-provider support
  • Composio Tool Router for intelligent tool routing across 300+ apps
  • Multi-step tool execution (up to 10 sequential tool calls per request)
  • MCP (Model Context Protocol) support for custom tool servers

Code Execution:

  • E2B Code Interpreter for sandboxed Python execution
  • 11 built-in skills: Excel/PowerPoint/PDF/Word generation, data visualization, web testing with Playwright, and more
  • Cloudflare R2 for generated file storage with signed URLs

Database & Auth:

  • Cloudflare D1 (SQLite) with Drizzle ORM
  • Clerk for user authentication
  • Per-user encrypted API key storage

Observability:

  • Arize Phoenix for LLM observability
  • OpenTelemetry for distributed tracing
  • Full visibility into token usage, latency, and multi-step execution flows

Challenges we ran into

The hardest part was tool orchestration at scale. With 300+ possible integrations, manually defining and routing tools would be unmanageable. The Composio Tool Router solved this by providing intelligent tool search and unified authentication across all connected services.

Other challenges:

  • Designing the multi-step execution flow so the model can use tool results to inform subsequent calls
  • Managing project-scoped skills so different projects can have different capabilities without "skill bloat"
  • Building a hybrid tool system that merges Composio integrations, E2B skills, and MCP tools into a unified interface
  • Implementing proper observability for debugging multi-step agentic loops

Accomplishments that I'm proud of

  • 300+ integrations available out of the box through Composio's Tool Router
  • Multi-step reasoning: The AI can chain up to 10 tool calls per request, using results from earlier steps to inform later decisions
  • Project-scoped skills: Each project can enable different capabilities, keeping the interface focused
  • Full observability stack: Every LLM call and tool execution is traced for debugging and optimization

What I learned

  • How the Composio Tool Router works and its power for solving the "too many tools" problem in agentic systems
  • Advanced Vercel AI SDK v6 patterns: multi-step execution, tool routers, and provider abstraction
  • The importance of observability in agentic systems—without tracing, debugging multi-step loops is nearly impossible
  • How MCP (Model Context Protocol) can standardize agent-tool interactions

What's next for Hive

Taking Hive from hackathon project to production:

  • Multi-user collaboration: Shared conversations where multiple users can interact with the same AI and tools
  • Custom skill creation: Let users define their own E2B skills with custom Python code
  • Workflow automation: Save and replay common multi-step workflows
  • Better context management: Improved handling of large tool results and conversation history

Built entirely in TRAE❤️

Built With

Share this project:

Updates