Consensus

Inspiration - AI agents are becoming increasingly capable of researching, analyzing, and making recommendations, but important decisions still require human judgment. We wanted to build a workspace where humans and AI could work together on the same decision instead of the AI simply giving a final answer.

Consensus was inspired by the idea of making AI an active collaborator while keeping humans in control of important decisions.

What it does - Consensus is a shared decision and research workspace for human + AI collaboration.

In Decision Mode, AI agents can use WebMCP tools to read the board, add options, score options, flag conflicts, resolve flags, and propose rankings. Humans can independently score options, review AI reasoning, dispute evaluations, and lock the final decision.

Consensus also includes Research Mode for extracting claims, checking evidence, identifying contradictions, and building citation relationships.

All important actions are persisted and can be reviewed through activity history and provenance.

How we built it - The frontend is built with React and Vite, with WebMCP integrated through document.modelContext.registerTool(...).

The backend uses FastAPI and is deployed on Google Cloud Run. Firestore stores shared board state, evaluations, flags, and activity history. Vertex AI with Gemini 2.5 Flash provides AI-assisted analysis.

The WebMCP tools connect to the same real backend operations used by the application, so agent actions update persistent state rather than a simulated demo.

Challenges we ran into - A major challenge was deciding what AI agents should be allowed to do and what should remain under human control. We wanted the agent to be useful without allowing it to silently make consequential decisions.

We also had to make WebMCP tool registration reliable as the application changed between different modes and states.

Finally, connecting the React frontend, FastAPI backend, Firestore, Vertex AI, and WebMCP into one working deployed application required careful coordination.

Accomplishments that we're proud of - We are proud that Consensus is a real deployed WebMCP application with persistent state, rather than just a prototype or mock demonstration.

The live application exposes six Decision-mode WebMCP tools that can interact with the shared workspace.

We are especially proud of the human-agent boundary: agents can investigate, evaluate, flag issues, and recommend actions, while humans retain control over scoring, overrides, and locking the final decision.

Our backend verification tests also pass checks for persistence, provenance, conflict handling, evaluation history, overrides, and locking.

What we learned - We learned that WebMCP becomes much more powerful when tools operate on real shared application state.

We also learned that human-AI collaboration is not only about giving agents more capabilities. It is equally important to define the actions they should not control.

Building Consensus also taught us how browser-native agent tools can work together with a persistent cloud backend and an AI reasoning layer.

What's next for Consensus - Next, we want to expand Consensus into more complex collaborative workflows.

We plan to explore multi-agent debate, stronger evidence and citation verification, configurable decision rubrics, richer provenance, and approval workflows with different levels of human confirmation.

Our long-term goal is to create a workspace where humans and AI agents can work together on complex decisions while keeping the reasoning, evidence, actions, and final authority visible and auditable.

Built With

• firestore
• google-cloud
• python
• react
• webmcp