Prisma: an AI workspace across desktop, web and MCP

The problem

Working with an AI assistant often means moving between a chat, an inbox and a calendar. The conversation can describe a task, but completing it requires access to the right tools, user context and connected accounts.

Prisma brings those pieces into a shared workspace. Its interface connects to an agent backend that manages conversations and tool execution, while MCP services expose capabilities from external applications.

My contribution

I built Prisma across the product interface, agent backend and service integrations. The project connects a React workspace with an Electron desktop application and a TypeScript Core that coordinates models, context and tools.

A central part of the work was defining boundaries between the client, the agent runtime and connected services. The same backend supports the desktop and browser experiences, while integrations can evolve behind a common tool interface.

Engineering decisions

Desktop and web share a React codebase. Vite runs the browser experience and Electron packages the desktop application. This keeps the workspace interaction consistent while allowing each environment to handle its own authentication flow.

Prisma Core owns agent state, context management and tool orchestration. The agent loop calls tools, incorporates their results and continues the conversation. Execution limits and failure handling give the runtime a way to stop and report when a connected service cannot complete an operation.

The GSuite MCP service exposes Gmail and Google Calendar tools, including email retrieval, calendar search, availability and meeting scheduling. Core handles OAuth and token management; the MCP service obtains access through an internal broker instead of duplicating the account connection flow.

Keeping the integration layer separate lets the application reuse its agent workspace while adding or changing external tools. Supabase provides the project's database layer, with migration and deployment documentation alongside the implementation.

What shipped

An implemented workspace spanning a desktop and web client, a Node.js agent Core and a dedicated MCP service for Google applications. The project demo shows the interface and the interaction model.

Prisma brought together the parts of an AI product that need to work as a system: a usable interface, an agent runtime, connected tools and account authorization. The result is a concrete foundation for exploring how an assistant can operate across everyday applications.

Architecture

Workspace interface

A shared React UI runs in the browser and in Electron on desktop.

Prisma Core

Coordinates conversation state, model responses, context and tool execution.

MCP integrations

Expose Gmail and Google Calendar capabilities through a standard tool interface.

Account connections

Core manages OAuth grants and brokers access for connected services.

TypeScript · React · Electron · Node.js · MCP · OAuth · Supabase

Explore the project

Explore the project through its demo and available resources.

• Watch the Prisma demo