/data/.
All operations are backed by real files on disk. Path resolution validates
against the root boundary to prevent directory traversal escapes.
Config
DiskVFS(root=..., folder_versions=True) takes the root path of the
directory to mount. Both READ and WRITE modes are supported.
VFS.md
The VFS.md an agent reads with Disk mounted at/<path>.
- Rendered
- Raw
root directory. For example,
if root="/srv/files" contains:
/data/ exposes:
../../etc/passwd are rejected - resolution is always confined
to the root boundary.
Cache
The Disk VFS usesIndexCacheStore with index_ttl = 60 (1 minute).
Directory listings are cached for up to 60 seconds before being refreshed
from disk. A configured index: replaces this lifetime with its own ttl, and the mount’s ttl: caps it.
Each listing is stored with its folder’s version: the folder’s device,
inode, change time and modified time. Under read: fresh, a later command
stats the folder and serves the cached listing while the version still matches, so
it pays one local stat per listed folder instead of a scan. A folder changed
in the last 2 seconds gets no version and is re-listed until it has been
quiet that long, since two changes inside one timestamp tick could leave its
times unmoved.
This assumes a local POSIX filesystem, where adding, renaming or removing an
entry moves the folder’s change time. A network or FUSE root (NFS, SMB,
rclone, s3fs, mirage’s own FUSE) may not, so turn the versions off there:
Example
Shell Commands
The Disk VFS supports the full set of shell commands since it operates on real file content (text, binary, JSON, CSV, etc.):Read Commands
Text Processing
File Operations
Path Utilities
Compression
Encoding
Use Cases
- Local directory access: Mount local directories for AI agents to read and process
- Sandboxed file access: Restrict agent file operations to a specific directory tree
- FUSE mounting: Expose disk files through a virtual FUSE mount for external tools
- Data pipelines: Process local datasets with shell-like commands
- Development: Test file operations against real data before deploying to cloud mounts