Role: Lead Product Designer
Team: Founding Machine Learning Developers, Product Manager, Frontend Engineers

Model Communication Protocol (MCP) is a framework for orchestrating multiple language models as coordinated agents—each with a defined role, scoped context, and turn in the reasoning chain. It shifts prompting from monolithic to modular: models summarize, critique, and rewrite each other’s outputs in sequence, forming a structured dialogue.

While MCP introduces a powerful mental model, current workflows are often fragmented—spread across notebooks, orchestration libraries, and opaque API calls. 

This project visualizes MCP from an interface perspective—making agent interactions transparent, inspectable, and user-directed. Users assign roles, define execution order, and trace how ideas evolve across model handoffs. Designed through both a product and engineering lens, the system supports reproducibility, orchestration, and step-level debugging.

Unblocking the Workflow

MCPs involve multiple moving parts — developers define specs, PMs scope features, engineers implement, and designers shape behavior. But without a shared interface, the flow breaks. Specs live across Notion, Slack, and code. This tool restructures that journey: model behavior is visualized, editable, and versioned — so every role stays in sync.

Designing for Dialogue

Agents can be confusing-- make model-to-model collaboration legible. I leaned on conversation as a UI structure — each agent speaks, critiques, or rewrites. Users assign roles like Summarizer, Critic, or Rewriter to selected models. The interface supports multi-step task orchestration through simple dropdowns and a guided prompt builder.

Conversation Playback

Outputs are presented as threaded messages, reflecting the sequence and evolution of ideas. The interface supports user feedback mid-dialogue, offering opportunities to intervene, redirect, or co-create.

Agent & Model Onboarding

Educational overlays help non-technical users understand how agent roles function, and how models differ in tone, reliability, and application.

 Session History & Sharing

A lightweight session dashboard where past conversations can be reviewed, duplicated, or exported. Each session displays a timestamp, the assigned models and roles, and a preview of the final output. Users can sort by agent, task type, or date to surface relevant collaborations.

Developer Console: Multi-Agent Config & Execution

This screen bridges interface design with the realities of modern LLMOps. It allows developers to structure multi-agent chains by assigning roles (e.g., Summarizer, Critic) to specific models, with full control over system prompts, temperature, and token limits. Configs are output as JSON payloads — not as an afterthought, but as a first-class asset for versioning and API execution.

Each agent’s response is logged with its inputs, latency, and token usage visible — because understanding model behavior at the step level is essential when chaining reasoning tasks across systems. Every interaction is replayable and forkable, supporting fast iteration and fine-grained debugging.

The API panel integrates directly with live endpoints and code exports, supporting transition from prototype to production. By exposing telemetry (rate limits, response times, token consumption) alongside structured configuration, this interface doesn’t just make LLM workflows usable — it makes them observable and maintainable.

Making Model Communication Legible

Building this interface began as an exploration of how multiple AI agents could collaborate more transparently—but it quickly evolved into a deeper question of how humans, too, might better understand, debug, and direct these interactions. What emerged is a system that treats multi-agent workflows not as code-first automations, but as legible, structured conversations.

Through role assignment, sequential reasoning, and step-level traceability, this tool reframes prompting as orchestration—making LLM behavior both observable and controllable.

Reflection & Future Directions