GHOSTBOUND: Open-Source Tool Gives ChatGPT a Verified View of Private Workspaces

TheFreakinOG · reddit · 2026-09-22

A Reddit developer open-sourced GHOSTBOUND, addressing a privacy boundary problem in AI coding workflows: he didn't want to hand his entire private workspace to a remote tool, nor let an "AI context optimizer" decide what the model sees.

The core idea: users define an explicit policy (View) over a Git workspace. GHOSTBOUND reads only allowed files from an exact Git commit (not the dirty working tree), creates a separate verified mirror with provenance records, and can optionally publish the View to Git. Stale views never auto-sync — updates must be explicit, an intentional security choice.

His actual workflow: ChatGPT/Sol acts as planner/reviewer over the verified View and produces a detailed /goal; Codex/Luna runs locally with real source code, validates assumptions, implements, and runs tests. The split also saves money: heavy thinking burns regular ChatGPT quota while Codex quota goes only to implementation.

v0.2 adds reusable Views, CURRENT/STALE/INVALID/UNVERIFIABLE status, disclosure deltas (what becomes newly exposed/changed/removed), and optional verified Git publication. The author migrated his real workflow to it after dogfooding, shadow-production updates, independent review, remote-consumer verification, and Gitleaks scanning.

Original post →

More from coding & agent

coding & agent channel →