cavi-ai/
GitHub ↗

Command line

The mlx-agent CLI is the provider-neutral core every adapter wraps. Standard library only; JSON envelopes for agents, human output for people.

Discovery and verification#

CommandPurpose
discoverDiscover MLX models for this host.
inspect-hostHost and runtime inventory without discovery.
adopt start/resume/statusResumable verify-and-recommend workflow.
bench run/aggregateMeasure a served model; aggregate export files.

Configuration and operation#

CommandPurpose
wire render/apply/status/rollbackConfirmation-gated config transactions.
serve start/stop/statusConfirmation-gated server lifecycle; --launchd for agents.
fleet render/applyOne-shot per-role router config.
doctor modelsInventory, drift, endpoint health; --prune for broken snapshots.
watch snapshot/diffStateful owned-model Hub digest.

Production#

CommandPurpose
convert scanRead-only GGUF inventory: converted, pending, and redundant copies.
convert start/statusConfirmation-gated quantization (4/8bit) from a cached repo or a GGUF file.
lora start/statusConfirmation-gated LoRA training.
fuse start/statusConfirmation-gated adapter fusion.

Research and design#

CommandPurpose
researchRead-only domain research packs (markdown + JSON).
blueprintGuidance-only project design packs.

Provider lifecycle#

providers, install, update, uninstall, and doctor manage the provider adapters themselves — all preview-then-confirm, all receipt-owned.

Every mutating command follows the same contract: preview without --confirm exits 2 with a diff and hash; --confirm --preview-hash applies exactly the reviewed plan; receipts make every mutation inspectable and reversible (except prune, which is loudly irreversible).