New in mi v1.7.5: "--sandbox" runs the agent inside Docker with your current directory mounted, to contain the blast radious in case of unexpected attacks.

mi Works with any OpenAI-compatible model (including local ones). Has REPL, skills, and recursive agents and only 30 lines of code, optimised for local models.

--- mi --sandbox -p 'fix tests until all are passing' ---