Continuum
A project management system built around persistent context and an MCP server for AI assistants.
At the beginning of 2026, I started using Notion to organise my growing portfolio of professional and personal projects. It gave me a central place for tasks, notes and project information, but after several months I was still struggling to turn that backlog into a realistic working schedule.
As a Product Manager working across multiple projects, context switching is a constant challenge. My tasks were in Notion, my time was managed in Google Calendar, and every new conversation with an AI assistant required me to explain the same project context again. I built Continuum to manage my projects and allocate time to them in Google Calendar from the same application, while giving AI assistants access to that shared project information through MCP.
Product screens and details
Cockpit
Today's calendar sessions, what is due and overdue, and what needs attention.
Board
Cards grouped by stage and ordered by date, with the next to work on at the top.
Card
A card holds the description, the calendar sessions booked for it, a checklist and linked cards.
Search
Search every card by title, project or status.
Select a screen to enlarge it. Screens are from the demo, which uses fictional data.
The story behind the project
The product
Continuum is built around a custom Model Context Protocol (MCP) server. The MCP interface allows an AI assistant such as Claude to look up projects, read the context and notes attached to cards, and create or update work through structured tools. Instead of keeping the status of a project inside a particular chat, the database holds the current state. I can start a new conversation, retrieve the relevant project and continue working from where I left off. The browser interface and AI assistant are two ways of working with the same underlying information.
Projects are organised into cards containing task descriptions, notes, status and supporting context. Work scheduled in the diary is represented as separate sessions, so one card can have several sessions without duplicating the task. The application connects to Google Calendar through its API to retrieve scheduled events and create calendar sessions for work on individual cards. This allows me to allocate time to a task and relate those sessions back to the project, rather than managing the backlog and diary independently. The browser interface uses TypeScript, Node.js and htmx to show the tasks and projects alongside the time allocated to them.
Projects, cards, sessions and their relationships are stored in SQLite. The application and MCP server run on a Google Cloud virtual machine using Debian Linux, with Caddy handling HTTPS requests to the Node.js service. My private instance has a passphrase login and daily SQLite snapshot backups. A separate demo environment allows visitors to explore the interface without accessing my working project data.
Building it with AI
I had two clear requirements from the outset, an MCP server that would let Claude read and update persistent project information and a direct integration with Google Calendar so I could read my diary and create sessions for the work I needed to do. Scheduling time is how I organise my working day, so being able to move between a project card and the calendar was fundamental to the product. Beyond those requirements, I wasn't entirely sure which features or workflows I would need. I decided to build a working version and refine it through daily use rather than try to specify everything upfront.
I worked through the initial architecture with Claude and ChatGPT, then imported approximately 196 cards from Notion so I could test it against real projects rather than sample data. As I used the application, I found gaps in how cards, sessions and project information were organised and added or changed functionality to address them. Some approaches were discarded when they didn't work in practice. The product evolved through repeated use, testing and changes to the implementation, while retaining the MCP and Google Calendar integrations as its core requirements.
I deployed Continuum to a Debian server on Google Cloud, with Node.js running the application, SQLite storing the data and Caddy handling HTTPS. I set up the deployment process, backups and a separate demo environment. There are 495 automated tests, which I use to check that changes haven't broken existing functionality.
Continuum is still being developed in this way. When I need a capability that the current version doesn't support, I assess the problem, build a bounded improvement and test it through my normal workflow. I haven't treated the initial release as a complete specification of what the product should eventually become.
Results and reflections
This is a working daily-use MVP, not a polished commercial release. The main result is a tool that supports my own planning and a practical testbed for AI agents working with structured operational data. The next improvements should follow observed use rather than speculative features. Using it every day gives me a direct feedback loop. I can see when the interface gets in the way of scheduling or when a conversation with Claude lacks the right context, and improve that specific journey rather than redesigning the entire tool.
The more transferable lesson is about connecting AI to dependable operational data. An assistant becomes more useful when its actions are grounded in a defined task and calendar model, but that also makes permissions, boundaries and reliability increasingly important.
Continue exploring
Interested in how I built it?
The full story covers the initial build, the problems I encountered and what I learnt.
Have an idea or a process you want working?
I take ideas from spec to a working product, and the same approach applies to automation and AI projects inside organisations.





