mcpd is a way to expose sandbox environments as MCP servers using just configuration. In the repo, I have some example config files that mimic popular coding harness tools.
I've been using it to give my cloud agents code search tools on private repos I manage. I have a script that syncs project code to the sandbox on deploy and I can connect it to my cloud agents and debug in Slack. It's awesome in combination with the Sentry MCP. I'm working on a secure way to have the agent open up PRs too.
My general thesis is the agent loop should be unprivileged and run remotely and tool execution should be separate and sandboxed. Disconnecting the loop and the tool execution environments opens up a lot of cool possibilities.
Comments URL: [https://news.ycombinator.com/item?id=49433211](https://news.ycombinator.com/item?id=49433211)
Points: 1
source & further reading
github.com — original article
Show HN: Sktime-CLI A command line tool to do all time-series tasks
An open protocol for AI-mediated product discovery
genai.md