I build and dogfood open tools for human-agent work.
Your agent already has tools, memory, Skills, MCP servers and a dashboard or two.
It may just be missing an Endroit.
That means "a place" in French. Yes, the pun is load-bearing.
I built one. It got out of hand.
A file-based place layer for agentic work.
I use it across real repositories and sessions so Codex, Claude and other tools can work from the same owned map. Repositories and services keep their own authority.
While dogfooding Endroit, I found responsibilities that belonged to the model rather than one implementation. I published them as Open Workplace, an implementation-neutral proposal.
I use it to ask a practical question: after the session ends, who owns the context, decisions, methods and results?
Some moments in human-agent work need an explicit, inspectable operation.
- Think It Through helps me stay with an idea long enough to explore or challenge it.
- Work This Way sets visible rules for the current session.
- Reality Check checks sources and assumptions without replacing the result.
- HACP is the draft protocol behind the Cards.
Open Workplace remains a proposal. Endroit is alpha and HACP is a draft. I publish the work while I dogfood it because awkward setups expose weak boundaries.