albedo

small gleam coding and persistent agent daemon with a detachable cli and a canvas ui albedo.sh

115

Fork 23

permissions modes + sandboxing #246

Open

aly.codes 1d ago

Open opened by

aly.codes 1d ago

being able to limit a session to a specific folder rather than it getting automatic full access would be very nice. along with this, being able to only allow specific tools with configurable permissions (like read-only, ask for approval, full access, and auto mode*) would be extremely nice

*for claude/codex, i'm not sure if their auto modes are able to be used outside of their harness or not, but at least Some classifier type "auto approve" thing would be nice

okami.mom 1d ago

yup, i’m working on a cpu classifier for this and will also do static checks as well

okami.mom 1d ago

only codex gives out a free classifier to use

okami.mom 1d ago

tbh me getting a model i really like would take a while so im more favorable in just detecting if you have docker or something, and putting the py kernel in there with a copy of the worktree with letting you choose what hosts it can ssh to and what not. i see tool gating as weird security theater thats annoying for both the agent and the user rather than meaningful gates that protects the agent from itself