Personal project · Early prototype · 2026
A project management agent that drafts the paperwork and shows where things stand
Describe a project in chat, and the agent drafts the charter, risk register, and stakeholder register from a library of templates. Anything it can't support from what you said gets flagged for you to confirm, never filled in. A dashboard then shows where the project stands.

What I built
Most of a project manager's paperwork follows known templates, and most of its numbers follow known formulas. So the agent's job is narrow on purpose: fill templates from what you actually said, call real calculators for the math, and say plainly what it doesn't know.
- One tool layer. Every capability is a function served over MCP, so it works from Claude Desktop or Claude Code, and the built-in chat calls the same functions in-process.
- Knowledge is a folder of Markdown files the agent searches and reads. One file is its rulebook and another holds the templates. Swap the folder and it follows a different method.
- An operating rule against invention: names, dates, budgets, and approvals the user didn't give are marked to confirm, and the dashboard counts what's still waiting on someone.
- Calculators for earned value, three-point estimates, critical path, and expected monetary value, tested against worked examples.
- One data model (risks, stakeholders, issues, milestones, change requests, lessons, and documents) that creates the Notion databases, validates what the agent writes, and sets the dashboard's columns.
- Storage in Notion, so records can be edited there and the dashboard reflects it, or in a local file for trying it out with no accounts.
Worth knowing
It's an early prototype. The tests cover the tools, the calculators, and the API, and the dashboard runs on my server. The live chat and the Notion connection haven't been run with real keys yet, and there's no login, so it stays on a private network.
I built it with Claude. I designed how it works, decided what it's allowed to do on its own, and tested it, and I'm the one accountable for what it does.
