Logixian Kitchen
An experiment in shared context for a team working with AI agents.
Creator and primary implementer · Five-person team
FastAPI · MCP · Web workspace
The coordination problem
Our Studio team was building a retirement-plan compliance platform while still learning the domain. Requirements, client feedback, and design decisions were spread across meetings, Confluence, Jira, and repositories. Each teammate’s agent saw a different slice of that project.
Generating more code or documentation was getting easier. Keeping that work aligned with the team’s current understanding was not. I built Kitchen to explore whether shared project access and explicit review boundaries could help, without requiring everyone to use the same model or editor.
What I tried
I built and deployed a web workspace and MCP server over the same provider layer. Teammates and their local agents can reach project context through a shared entry point. Jira, Confluence, and GitHub remain authoritative, rather than becoming copies inside another project database.
For generated tickets and documents, Kitchen stages a proposal for a person to revise, reject, or publish. Shared recipes describe review criteria rather than prescribing every agent step. The intent is to make contributions reviewable against team requirements while leaving the agent’s research and execution flexible.
Context-access traces offer another angle: which resources did an agent retrieve before proposing a change? That can help inspect missing context, but retrieval alone does not show that the agent understood or used it.
Does a team need this layer?
I’m still not sure. Giving an agent more context and a larger token budget often gets the job done. The team can also keep working directly in its existing tools without Kitchen. A shared context layer adds its own maintenance work and another place for guidance to become stale.
The question is whether it makes enough difference to justify that cost. I want to compare the same tasks with and without prepared context, looking at review effort, missed requirements, and token use. Kitchen is a working experiment, not yet evidence that every agent-assisted team needs a context layer.