How the daily surfaces work across four tabs - Search, Library, Settings, and Help: the search-and-install flow, the library that manages what the machine already holds, and the settings that configure the manager.
Search mode queries the remote catalog and lists one row per model, with its
quantization, size, parameter count, and context length. ↑/↓ selects,
Enter installs.
An install resolves the remote file, downloads it, and verifies its SHA-256 against the digest the catalog advertised. A model whose upstream advertises no checksum still installs, but is reported as unproven rather than verified: the manager reports what it can prove instead of assuming success.
Progress is shown while bytes are in flight; Esc returns to wherever the
install started.
Tab and Shift+Tab cycle between the four tabs (Search, Library, Settings,
Help). You can also jump straight to one with Alt+1 (Search), Alt+2
(Library), Alt+3 (Settings), or Alt+4 (Help).
Library mode gives full control over the models already on this machine:
| Key | Action |
|---|---|
↑ / ↓ |
Move through the installed models |
i / Enter |
Inspect one model: revision, exact path, size, full digest |
v |
Prove the replica against its recorded digest, or the remote's advertised checksum when none is recorded |
d |
Delete a model, after a y/n confirmation |
p |
Prune leftovers: orphan .sha256 / .tags notes and emptied directories |
r |
Re-read the library from disk |
h |
Open help |
Esc |
Close a popup, else return to the model table |
Ctrl+C / Ctrl+Q |
Quit |
The header reports where the library lives, how many models it holds, how much space they take, and how many are verified or broken.
- verified - the library recorded a digest it proved for these bytes.
- unproven - the bytes are on disk but nothing has proved them, because no
checksum is recorded here and none is advertised upstream.
vproves a replica when a checksum exists in either place; without one it stays unproven. - broken - the bytes no longer match the digest recorded for them.
The orb in the film corrupted everyone who took it on faith. v replaces faith
with arithmetic.
Listing the library never hashes a file, which is what keeps it fast with
multi-gigabyte models. Proving bytes is v's job, on demand, for one model.
Deleting removes the replica, its .sha256 digest note, its .tags note, and
the directories that model alone needed - never the library root.
Pruning discards only what stands for no model: orphan .sha256 or .tags
notes whose replica is gone, and directories left holding nothing. A model you
installed is never a leftover, proven or not, and a file the manager did not put
there is never touched.
Taarna struck only what had earned it. So does p.
The Settings tab (Alt+3) edits the operator's persisted configuration. It
shows four fields:
- Hugging Face API Token - masked as
*except while being edited. - Hugging Face Endpoint - the catalog base URL.
- Hugging Face Cache Directory - where downloads are staged.
- Model Download Path - the library root models are installed into.
Controls: ↑/↓ select a field, Enter edits the selected field, Esc
cancels the current edit, and s saves and applies the settings.
Instructions for testing and validating models downloaded by LocalNar with
llama-cli, llama-server, and llama-bench across different hardware
configurations are documented in
docs/model-validation.md.