Codex Cache
Codex stores a local cache under ~/.codex/cache for app and connector metadata that speeds up startup, workspace discovery, provider lookups, model lists, icon and image fetches, and rendered UI resources. Over time this cache can accumulate stale JSON metadata, HTTP response bodies, image thumbnails, compressed asset files, and Electron Chromium disk-cache entries that no longer match the current app version or remote connector state. Kudu removes only those disposable cache files so Codex can rebuild fresh metadata without touching sessions, archives, memories, attachments, generated images, credentials, configuration, plugins, or installed runtimes.
Why clean Codex Cache?
- Stale connector metadata leaves renamed or removed providers in the picker, so Codex shows options that fail immediately when you try to connect
- Cached model and capability manifests can lag behind the server, causing missing models, wrong context limits, or disabled tools until the metadata is refreshed
- Corrupted Electron HTTP cache entries make parts of the interface load as blank panels or broken icons even though your account and settings are intact
- An oversized cache directory slows startup and workspace indexing because Codex spends extra time scanning old metadata and expired response files
- Outdated app asset bundles in the disk cache can leave mismatched UI text or controls after an update, with buttons present but actions failing until fresh resources are downloaded
- Stale image thumbnails and preview artifacts can make connector cards or recent items display the wrong artwork, then suddenly correct themselves after a manual refresh
- If Codex stores cache indexes in SQLite, page fragmentation can waste disk space and slow lookups; VACUUM rewrites the database into a compact file without deleting actual user rows
Cache paths Kudu targets
Windows
~/.codex/cache |
macOS
~/.codex/cache |
Linux
~/.codex/cache |
Common questions about Codex Cache
Download Kudu and reclaim your disk space.
Available on Windows, macOS, and Linux. No account required, no feature gates, no telemetry without consent. All cleaning targets are open source and community-auditable.