Problem
A chat assistant can plan code changes but cannot run them. Unlimited shell access to a personal machine is not safe.
Value
ChatGPT gets access to code, files, the command line and tests only inside allowed projects, and every action is logged.
Role
Server and security design, directing the work. Agents wrote most of the code.
- Over 40 tools: version control, files, command line, reading analytics data, local language models and long agent runs.
- Security: a list of allowed projects, read-only folders, a list of blocked commands, sign-in with a token, connections from the local machine only and a log of every action.
- Tool output is capped so long results don’t fill up the model’s memory.
- Access to local skill libraries (Agent Skills), so several agents can use the same skills.
Limits
- A single-user tool for one machine, not a hosted service.
Sources
- Latest change
- 88 code changes, 29 Jul–3 Oct 2026