I'm Ashutosh, the maintainer. I spent 103 hours building the initial toolkit around a practical need: connecting the source a developer inspected to the exact change and command they approved. Agentic Workbench has a Python CLI/SDK, eight role routes, versioned skills, saved tasks, source citations, and reviewed change/rollback workflows. A typed Node adapter and restricted MCP server use the same policy implementation. Start with the offline init, scan and explain workflow in the README. No API key is needed for that trial. Optional OpenAI use requires your own credentials and an explicit network grant. This is a pre-1.0 developer preview. Guarded mutations and execution require POSIX; Windows is limited to the documented read-only subset. Trusted-host execution is weaker than Docker isolation and requires explicit authority. 0.1.0 is the published package. 0.1.1 is a prepared candidate, not a registry release. I'd value technical feedback on the source context, approval flow and installation experience.

Agentic Workbench

Supervised coding agents with reviewable changes

Supervised coding agents with reviewable changes