I'll often work with Claude on a project — currently making a puzzle where a little robot follows cards kids put down — and want some form of project management beyond the context window.

As in, Claude will present a great plan (eg: buy these 12 components, etc) and some of them will be for Claude and often some of them will be for me.

I don't find plan mode, project spaces nor context windows to be sufficient at project management — they too often forget nuances to the point that I can't really just say "remind me of the components I need to buy" and expect a great response right away.

At the same time, I am also interested in scaling my token usage (token maxxing, as I like to call it) by running agents or even agent teams in parallel. I find the Claude plan mode to be okay for this, but I still struggle to organize a backlog to the point that a team of agents can work through it.

So, I made Limoncello which is Trello for agents.

Just like Caldave, agents can sign up and use this service without human intervention. No email verification or anything else, agents can simply curl an endpoint and get access.

As part of the creation, Limoncello will suggest that the agent ask their user to install the Limoncello MCP — which does all the things you expect:

• Create a board
• Create a ticket
• Tag a ticket
• Move tickets around, etc.

Making agents use the damn thing

When an agent creates an API key it is then asked to install the MCP. After installing the MCP there is an limoncello_onboard tool that asks to be ran once per project.

It basically tells the agent to create a board (using the board create tool) and then store that data in the most canonical documentation for that project (Claude.MD for example)

Then — and I am quite proud of this — it adds hooks to the projects settings.json:

• After ExitPlanMode: Prompt to create Limoncello cards for non-trivial plans (tracks each step)
• After TaskCompleted: Check for matching Limoncello cards and move them to done, then commit changes.