CLI reference

Generated at build time from the CLI command table and help. These are the current commands on main. Released versions can differ; use the installed command’s --help.

Global options

Usage: mlx-bun [options]

An option-first invocation starts the server; use the serve options below.

Option	Description
-h, --help	Show help
-v, --version	Show version

serve

Serve a local model with continuous batching and the web app

Usage: mlx-bun serve [query] [options]

Option	Description
--model <value>	Model directory or cached registry query (overrides positional query)
--query <value>	Cached model query when no positional/model override is supplied
--host <value>	Bind address [default: 127.0.0.1]
--port <value>	Listen port; 0 chooses a free port [default: 8080]
--batch <value>	Continuous batching capacity, including at 1 [default: 8]
--decode-concurrency <value>	Alias for —batch (main’s spelling); the same continuous capacity
--max-tokens <value>	Default completion cap when a request omits one
--thinking <value>	Default thinking mode: on | off; requests may override
--temperature <value>	Default sampling temperature [0..5]
--temp <value>	Alias for —temperature
--top-p <value>	Default nucleus sampling [0..1]
--top-k <value>	Default top-k sampling
--kv-quant <value>	KV quantization: off | config | 4 | 8 | turbo[:kv] (turbo is k8v3); config without the model’s kv_config.json stays bf16, and a scheme the model’s cache cannot take is refused at startup [default: off]
--l1	Numerical preset: KV off, unfused SDPA (the default); explicit KV/kernel flags override
--l2	Numerical preset: model-config KV, fused SDPA; wins over —l1, explicit KV/kernel flags override
--fused-sdpa <value>	Startup SDPA override: on | off [default: on for KV config, off otherwise]
--kv-budget <value>	Aggregate batch KV budget, decimal GB; unset means no budget
--prompt-cache <value>	RAM prompt cache cap, GiB; 0 disables [default: 8 GB]
--ssd-cache <value>	Saved prompt/KV state directory, or off to disable it [default: MLX_BUN_HOME/kv]; one directory per model identity, so a model returns to its saved state
--ssd-cache-max <value>	Byte budget for the whole saved-state directory across every model, GiB; 0 means unlimited [default: 20]
--ssd-cache-verify	Verify tensor hashes on SSD restores
--ssd-demote-idle <value>	Idle seconds before SSD demotion; 0 disables [default: 300]
--generation-checkpoint <value>	Checkpoint every N generated tokens into the saved state
--memory-budget <value>	Admission-control memory budget, decimal GB; requests that cannot fit are rejected instead of crashing the GPU
--model-budget <value>	What all resident models may use together, decimal GB: a model that fits loads beside the others, otherwise the least recently used one is drained, its saved state flushed, and released first [default: 70% of the GPU’s recommended working set]
--context-length <value>	Context tokens a memory-planning runtime reserves [default: its preset]; models without a plan ignore it
--force-wire	Wire weights into memory at load
--expert-offload	MoE only: serve experts from a page-aligned file mmap built on first use
--allow-private-media	Let image_url/audio_url parts fetch from private, loopback, and link-local hosts
--hlg-sampling <value>	Piecewise tone-curve (HLG) sampling: on | off [default: off]
--hlg-width <value>	HLG mid-region half-width, nats [default: 4]
--hlg-shoulder <value>	HLG highlight rolloff scale, nats [default: 4]
--hlg-toe <value>	HLG shadow rolloff scale, nats [default: 6]
--hlg-pivot-offset <value>	HLG pivot: nats below the top token [default: 6]
--draft-model <value>	Speculative decoding draft: a path or cached query resolved like the main model; kind auto-detected
--draft-kind <value>	Draft kind override: two-model | assistant | dspark | deepspec | mtp | ngram (ngram is model-free; mtp alone mounts /mtp/)
--num-draft-tokens <value>	Drafts per verify round: an integer >= 1, or adaptive (each round’s count, from none up to the drafter’s width or the width the graph measured for its kind, from recent acceptance and the graph’s measured verify costs; needs a graph measured on this machine) [default: 3; ngram: 10]
--ngram-max <value>	Prompt-lookup longest k-gram, integer >= 1 (ngram only) [default: 3]
--ngram-min <value>	Prompt-lookup shortest k-gram, integer >= 1 (ngram only) [default: 1]
--mtp <value>	Checkpoint-native multi-token-prediction draft head: on | off [default: on]; models without one ignore it
--paged-kv	Paged KV cache, composed when the model loads (env mirror MLX_BUN_PAGED_KV=1): every request keeps its full-attention keys and values in fixed blocks, stored as —kv-quant names for each layer
--paged-kv-block-size <value>	Tokens per KV block with —paged-kv, integer >= 1 [default: 256]
--paged-kv-direct	With —paged-kv, read at most eight queries (decode, short prefill and verification windows) with the direct Metal kernel over the blocks (Lab numerics) instead of gathered blocks through the fused SDPA; TurboQuant pages and affine widths that straddle a 32-bit word read gathered
--adapter <value>	Mount a LoRA adapter directory at startup as the default for requests without an adapter field
--adapter-path <value>	Alias for —adapter (mlx_lm spelling)
--whisper-model <value>	Whisper checkpoint (path or cached query) for /v1/audio/* beside the chat model [default: the first downloaded whisper model, resolved on the first request]; serving a Whisper checkpoint ALONE starts a transcription-only server
--whisper-idle-unload <value>	Seconds the Whisper weights stay loaded after a take; 0 releases them right after every take [default: 0]
--whisper-resident	Never release the Whisper weights
--preload	Transcription-only server: load the Whisper weights before listening instead of on the first request
--no-open	Do not open the web app in an interactive terminal
--in-process	Load the models in this process instead of one worker process per resident model. Isolated serving is the default: this process keeps the app, web chat and jobs and loads no model, each worker is crash-isolated and respawns after a crash, and the workers are held by memory fit (—model-budget)
--isolate	Deprecated, accepted and ignored: isolated serving is the default (see —in-process)
-h, --help	Show help

Use mlx-bun serve --help for terminal help.

generate

Generate text once from a local model

Usage: mlx-bun generate [query] [prompt] [options]

Option	Description
--query <value>	Cached model query when no positional query is supplied
--prompt <value>	Prompt text (or second positional argument); - reads stdin
--system-prompt <value>	System message placed before the prompt in the chat template (ignored with —raw)
--raw	Skip the chat template and tokenize the prompt verbatim
--max-tokens <value>	Completion cap [default: 256]
--temperature <value>	Sampling temperature [default: 0]
--temp <value>	Alias for —temperature
--top-p <value>	Nucleus sampling
--top-k <value>	Top-k sampling
--min-p <value>	Min-p sampling: keep tokens with probability >= min-p times the top token’s [0..1]
--min-tokens-to-keep <value>	Tokens min-p never filters out, integer >= 1 [default: 1]
--xtc-probability <value>	Probability per step of XTC (exclude top choices) sampling [0..1]
--xtc-threshold <value>	Probability a token must exceed to be an XTC removal candidate [0..0.5]
--adapter <value>	Mount a LoRA adapter directory for this generation
--adapter-path <value>	Alias for —adapter (mlx_lm spelling)
--seed <value>	Sampler seed
--kv-quant <value>	KV quantization: off | config | 4 | 8 | turbo[:kv] (turbo is k8v3); config without the model’s kv_config.json stays bf16, and a scheme the model’s cache cannot take is refused at startup [default: off]
--quantized-kv-start <value>	Token count at which KV quantization starts with —kv-quant 4 | 8 | turbo | config [default: 0]
--l1	Numerical preset: KV off, unfused SDPA (the default); explicit KV/kernel flags override
--l2	Numerical preset: model-config KV, fused SDPA; wins over —l1, explicit KV/kernel flags override
--fused-sdpa <value>	Startup SDPA override: on | off [default: on for KV config, off otherwise]
-h, --help	Show help

Use mlx-bun generate --help for terminal help.

embed

Embed local text and print vectors

Usage: mlx-bun embed [query] [text] [options]

Option	Description
--query <value>	Cached model query; defaults to the first downloaded embedding model
--text <value>	Text to embed; otherwise second positional or nonempty stdin lines
--instruct <value>	Query instruction; omit for document embeddings
--json	Print one OpenAI-style embedding list instead of one vector per line
-h, --help	Show help

Use mlx-bun embed --help for terminal help.

perplexity

Perplexity of a model over a local text/JSONL dataset

Usage: mlx-bun perplexity [query] [options]

Option	Description
--model <value>	Model directory or cached registry query (mlx_lm spelling of the positional)
--query <value>	Cached model query when no positional is supplied
--data-path <value>	.jsonl ({“text”: …} rows) or plain .txt (required)
--sequence-length <value>	Tokens per row [default: 512]
--num-samples <value>	Rows to score; -1 = all available [default: 256]
--batch-size <value>	Rows per forward [default: 8]
--seed <value>	Sample-shuffle seed [default: 123]
-h, --help	Show help

Use mlx-bun perplexity --help for terminal help.

compile-ane

Compile the Neural Engine programs this Mac’s models use that the system cache lacks (the install step)

Usage: mlx-bun compile-ane [options]

Option	Description
--check	List each program and whether the compile cache holds it; compile nothing (exit 1 when one is missing)
--quiet	Print only the programs it compiles (nothing when every program is cached)
-h, --help	Show help

Use mlx-bun compile-ane --help for terminal help.

transcribe

Speech-to-text from an audio file with a local Whisper model (no server)

Usage: mlx-bun transcribe <audio-file> [query] [options]

Option	Description
--model <value>	Whisper model directory or cached query (overrides the positional query) [default: the first downloaded whisper checkpoint]
--query <value>	Cached Whisper query when no positional/model override is supplied
--language <value>	ISO code or name (en, japanese); auto or omitted detects from the first 30 s
--task <value>	transcribe | translate (to English) [default: transcribe]
--beam-size <value>	Beam search width [default: greedy]
--temperature <value>	One sampling temperature; omitted runs the (0, 0.2, …, 1.0) fallback ladder like mlx-whisper
--no-fallback	Temperature 0 only, no retry ladder
--prompt <value>	Initial prompt (vocabulary hints / style)
--no-timestamps	Decode text only (<|notimestamps|>)
--no-condition	Do not condition each window on the previous text
--vad	Silero VAD gate: no detected speech prints an empty result and Whisper never runs (weights: ggml-org/whisper-vad in the HF cache)
--vad-threshold <value>	Speech probability threshold for —vad [default: 0.5]
--vad-model <value>	Explicit Silero ggml path for —vad
--vad-trim	Accepted with —vad; the clip is still transcribed whole, as in main
--word-timestamps	Word-level timestamps (verbose_json words)
--faithful	Run the oracle-parity graph instead of the fast path
--audio-ctx <value>	Lab: encode only n of the 1500 encoder positions (whisper.cpp -ac); degrades below ~1024
--format <value>	text | json | verbose_json | srt | vtt [default: text]
--verbose	Print each segment with its time range as it decodes, then a realtime-factor summary (stderr)
-h, --help	Show help

Use mlx-bun transcribe --help for terminal help.

dictate

Push-to-talk: stream your microphone into Whisper, print/copy/type the text

Usage: mlx-bun dictate [query] [options]

Option	Description
--model <value>	Whisper model directory or cached query for in-process transcription (overrides the positional query) [default: the first downloaded whisper checkpoint]
--query <value>	Cached Whisper query when no positional/model override is supplied
--server <value>	Use a running mlx-bun server’s /v1/audio/sessions instead of loading the model here
--hotkey <value>	Hold a key to talk instead of Enter toggling: macOS virtual keycode, value optional [default: 61 = Right Option]; needs Input Monitoring for your terminal
--language <value>	Language code; auto detects [default: en]
--beam-size <value>	Beam search width [default: greedy]
--prompt <value>	Initial prompt
--vocabulary <value>	Comma-separated hint terms, fitted into the prompt budget
--no-vad	Skip the Silero gate (by default silence never runs Whisper)
--idle-unload <value>	Release the weights after this many idle seconds; 0 = right after each take [default: 30]
--resident	Keep the weights loaded
--copy	Copy the transcript to the clipboard (pbcopy)
--type	Type it into the frontmost app via System Events keystrokes (needs Accessibility for your terminal)
--type-delay <value>	Seconds to wait before typing so you can Cmd-Tab [default: 1 in Enter mode, 0 with —hotkey]
-h, --help	Show help

Use mlx-bun dictate --help for terminal help.

convert

Quantize an HF model into a local MLX snapshot (mlx_lm.convert counterpart)

Usage: mlx-bun convert [repo-or-path] [options]

Option	Description
--hf-path <value>	Source model: local path, downloaded model, or HF repo id (fetched first); —model and the positional are aliases
--model <value>	Alias for —hf-path
--mlx-path <value>	Output directory; must not already exist [default: ~/.mlx-bun/models/-bit]
-q, --quantize	Quantize the model (uniform affine); without -q or —target-bpw the model is rewritten with —dtype and/or -d only
--q-bits <value>	Bits per weight: 4 or 8 [default: 4]
--q-group-size <value>	Quantization group size: 32 or 64 [default: 64]
--upload-repo <value>	Push the converted model to this Hugging Face repo afterwards (write token checked first)
--target-bpw <value>	Mixed precision target bits-per-weight, e.g. 4.5: OptiQ sensitivity sweep + per-layer knapsack; implies -q
--candidate-bits <value>	Comma list the knapsack may pick from [default: 4,8]
--calibration-mix <value>	“optiq” or a JSONL path [default: optiq]
--n-calibration <value>	Calibration samples [default: 2]
--rotate-weights	Fold the model’s offline TurboQuant rotation before quantization (auto-detects Llama/Qwen3.5/Qwen MTP)
--rotation-seed <value>	Deterministic rotation seed [default: 42]
--q-mode <value>	Quantization mode: affine, or trellis for packed Trellis (TCQ) MLP tensors of an HF-layout Qwen3.5-family model (rotation fold implied; implies -q) [default: affine]
--trellis-bits <value>	—q-mode trellis: bits per coded weight, 1-8 [default: 3]
--trellis-k-map <value>	—q-mode trellis: allocation JSON with per-tensor bits (budgets[—trellis-k-budget].kmap, optional affine_map)
--trellis-k-budget <value>	—q-mode trellis: budget key inside —trellis-k-map [default: 3.00]
--trellis-ldlq <value>	—q-mode trellis: directory of per-layer Hessian factors (layer-NNN-mlp|down.safetensors) enabling BlockLDLQ error feedback
--trellis-reuse <value>	—q-mode trellis: comma list of packed artifacts to copy unchanged-geometry tensors from instead of re-encoding
--trellis-down-axis <value>	—q-mode trellis: dim of down_proj the trellis runs along, out (rotated dim) or in [default: out]
--trellis-interleave	—q-mode trellis: reorder eligible 3-bit axis-0 codes into the two-block scatter layout
--trellis-layers <value>	—q-mode trellis: trellis-code only the first N decoder layers (the rest ride the 3-bit affine base tier)
--dtype <value>	Dtype of the non-quantized tensors and of the quantization scales/biases: float16 | bfloat16 | float32 [default: bf16 scales/biases, other tensors unchanged]
-d, --dequantize	Dequantize a quantized model to dense weights (not with -q)
--quant-predicate <value>	Not supported (mlx-lm’s mixed_* recipes need 2/3/6-bit); use —target-bpw for mixed precision
-h, --help	Show help

Use mlx-bun convert --help for terminal help.

train

Fine-tune a LoRA adapter on your data (sft | dpo | orpo)

Usage: mlx-bun train [model] [options]

Option	Description
--query <value>	Model to fine-tune when no positional query is supplied (auto-picks the default model if omitted)
--data <value>	Dataset dir with train.jsonl (+ optional valid.jsonl); rows are {prompt, chosen, rejected} for dpo/orpo, {messages|text} for sft (required)
--method <value>	sft | dpo | orpo [default: orpo]
--adapter <value>	Output adapter dir [default: ~/.mlx-bun/adapters/-]
--iters <value>	Training iterations [default: 100]
--lr <value>	Learning rate [default: orpo 1e-5 · dpo 5e-5 · sft 2e-4]
--rank <value>	LoRA rank [default: orpo 16 · else 8]
--scale <value>	LoRA scale [default: orpo 2.0 · else 1.0]
--seq <value>	Max sequence length [default: the model’s own; 4096 when it declares none]
--batch <value>	Batch size [default: 1]
--grad-accum <value>	Gradient accumulation steps (effective batch = batch × grad-accum at batch-size-1 memory) [default: 1]
--grad-clip <value>	Gradient-norm clip (0 = off) [default: 1.0]
--seed <value>	Data-shuffle / init seed [default: 0]
--val-size <value>	Max validation examples per eval [default: 256]
--lambda <value>	ORPO odds-ratio weight [default: 0.1]
--sft-scope <value>	ORPO chosen-NLL scope: full (paper/TRL-faithful, prompt+response) | response (pre-2026-07 runs, bit-exact) [default: full]
--seg <value>	Layers per segment (segmented backward; orpo default 2)
--save-every <value>	Crash-safe mountable checkpoint every n steps
--resume <value>	Warm-start LoRA weights from a checkpoint/adapter dir
--no-flash	Disable the flash-CCE Metal head (use the MLX fused head)
--no-prefix	Disable prefix-sharing (two-forward branches)
--no-segment	Disable the segmented backward (hold all activations)
--num-layers <value>	Layers to adapt, counted from the last; -1 = all [default: -1]
--steps-per-report <value>	Steps between loss reports [default: 1]
--steps-per-eval <value>	Steps between validations [default: every —save-every steps, else never]
--dropout <value>	LoRA dropout [default: 0]
--weight-decay <value>	AdamW weight decay [default: 0.01]
--grad-checkpoint	Recompute activations in the backward pass to save memory
--dry-run	Inspect the dataset + print the resolved plan, don’t train
-h, --help	Show help

Use mlx-bun train --help for terminal help.

draft

Produce a speculative-decoding drafter: regen | train | calibrate | quantize

Usage: mlx-bun draft <action> [model] [options]

Option	Description
--model <value>	Target model (regen, train, calibrate) or DeepSpec drafter directory (quantize); the option spelling of the second positional
--topics <value>	regen: text file with one topic per line the target writes an article about (required)
--data <value>	train: shard directory from regen (required) · calibrate: JSONL of {prompt} rows, a string or chat messages (required)
--out <value>	Output directory [default: regen ~/.mlx-bun/datasets/dspark- · train ~/.mlx-bun/models/-dspark · quantize ~/.mlx-bun/models/-affine-q-g · calibrate: in place]
--tap-layers <value>	regen/train: target layers the drafter reads, comma separated, the last normally the layer count (post-final-norm). MUST match between regen and train [default: 20,31,41,42 (gemma-4 e4b)]
--max-resp <value>	regen: generated tokens per topic [default: 320]
--seqs-per-shard <value>	regen: sequences per shard file [default: 32]
--min-resp <value>	regen: drop responses shorter than this [default: 6]
--iters <value>	train: optimizer steps [default: 6000]
--batch <value>	train: anchors per step [default: 8]
--gamma <value>	train: draft block length [default: 5]
--max-ctx <value>	train: cap on the prefix context each anchor attends [default: 512]
--lr <value>	train: peak learning rate [default: 1.5e-3]
--warmup <value>	train: warmup steps [default: 150]
--eval-every <value>	train: steps between held-out evaluations; the best held-out τ is what is saved [default: 500]
--eval-anchors <value>	train: anchors per held-out evaluation [default: 256]
--seed <value>	train: init and sampling seed [default: 0]
--d-draft <value>	train: drafter width [default: 1024]
--n-heads <value>	train: drafter attention heads [default: 8]
--layers <value>	train: drafter backbone depth [default: 5]
--markov-rank <value>	train: sequential-head rank [default: 256]
--seq-head <value>	train: sequential head, markov (Eq 5) | rnn (Eq 6) [default: markov]
--resume	train: warm-start weights from an existing checkpoint in —out (the optimizer and schedule restart)
--drafter <value>	calibrate: the trained drafter directory whose dspark.json receives the thresholds (required)
--n <value>	calibrate: prompts to run [default: 32]
--max-tokens <value>	calibrate: generated tokens per prompt [default: 128]
--precision <value>	calibrate: acceptance precision each kept position must reach, in (0, 1] [default: 0.5]
--min-samples <value>	calibrate: positions with fewer observations are never pruned [default: 50]
--force	calibrate: recalibrate a drafter that already carries thresholds
--bits <value>	quantize: 4 | 8 [default: 4]
--group-size <value>	quantize: 32 | 64 [default: 64]
-h, --help	Show help

Use mlx-bun draft --help for terminal help.

train-watch

Live dashboard for a training run (tails /metrics.jsonl)

Usage: mlx-bun train-watch [adapter-dir] [options]

Option	Description
--adapter <value>	Adapter directory to watch; accepted for the positional [default: the latest run in ~/.mlx-bun/adapters]
-h, --help	Show help

Use mlx-bun train-watch --help for terminal help.

fuse

Merge a LoRA adapter into the base weights (writes a standalone snapshot)

Usage: mlx-bun fuse [model] [options]

Option	Description
--model <value>	Base model (registry query or a snapshot path); the mlx_lm.fuse spelling of the positional
--adapter <value>	Adapter directory (adapters.safetensors + adapter_config.json) [default: adapters]
--adapter-path <value>	mlx_lm.fuse alias for —adapter
--save-path <value>	Output model directory [default: ~/.mlx-bun/models/-fused]
--dequantize	Write dense weights instead of the base’s quantized layout (drops the config’s quantization block)
--export-gguf	Not supported (mlx_lm.fuse flag); the command exits with an error
--gguf-path <value>	Not supported (mlx_lm.fuse flag); the command exits with an error
--upload-repo <value>	Push the fused model to this Hugging Face repo afterwards (write token checked first)
-h, --help	Show help

Use mlx-bun fuse --help for terminal help.

get

Download a model from Hugging Face (resumable, verified)

Usage: mlx-bun get <org/repo | substring> [options]

Option	Description
--revision <value>	Git revision [default: main]
-h, --help	Show help

Use mlx-bun get --help for terminal help.

ls

List downloaded models (one canonical revision per repo)

Usage: mlx-bun ls [query] [options]

Option	Description
--vision	Only vision-capable models
--max-size <value>	Filter by weight size, e.g. 10GB or 800MB
--all-revisions	Show each snapshot; canonical revision marked *
-h, --help	Show help

Use mlx-bun ls --help for terminal help.

scan

Re-index the Hugging Face cache without reading tensor bytes

Usage: mlx-bun scan [options]

Option	Description
-h, --help	Show help

Use mlx-bun scan --help for terminal help.

fit

Estimate model memory and decode speed on this machine

Usage: mlx-bun fit <query> [options]

Option	Description
--ctx <value>	Context tokens [default: 8192; a memory-planning runtime defaults to its own preset]
--kv-quant <value>	KV estimate: 4 | 8 | config | off [default: off]
--skus	Print the Apple Silicon SKU matrix
-h, --help	Show help

Use mlx-bun fit --help for terminal help.

gc

Reclaim superseded snapshots and dead blobs (preview by default)

Usage: mlx-bun gc [options]

Option	Description
--yes	Actually delete the planned snapshots and blobs
--dry-run	Never delete, even with —yes
--force	Also prune superseded snapshots with otherwise unique files
-h, --help	Show help

Use mlx-bun gc --help for terminal help.

upload

Push a local model directory to the Hugging Face Hub (mlx_lm.upload counterpart)

Usage: mlx-bun upload [options]

Option	Description
--path <value>	Local model directory to upload (required)
--upload-repo <value>	Hub repo id, org/name or bare name (required)
--private	Create the repo as private (mlx-bun extension)
-h, --help	Show help

Use mlx-bun upload --help for terminal help.

memory

Your local AI’s personal wiki: set it up, inspect it, run synthesis, schedule it

Usage: mlx-bun memory [subcommand] [args] [options]

Option	Description
--since <value>	synthesize: only conversations newer than this (parsed; the pipeline does not consume it yet)
--model <value>	synthesize: synthesis model override (parsed; reserved)
--dry-run	synthesize: plan the stages only, never write the vault
--limit <value>	Stage workers: cap the work processed this pass (segment, extract, synthesize-stage, link)
--convs <value>	Stage workers: comma-separated conversation ids to restrict the pass to
--at <value>	schedule: local wall-clock time for the nightly job, 24h HH:MM [default: 03:00]
--host <value>	Run the model calls on a serving mlx-bun at this host instead of loading the memory task model (127.0.0.1 when only —port is given)
--port <value>	Port of that server (8080 when only —host is given)
-h, --help	Show help

Use mlx-bun memory --help for terminal help.

setup

Set up your local AI’s memory wiki (alias of mlx-bun memory)

Usage: mlx-bun setup [subcommand] [args] [options]

Option	Description
--since <value>	synthesize: only conversations newer than this (parsed; the pipeline does not consume it yet)
--model <value>	synthesize: synthesis model override (parsed; reserved)
--dry-run	synthesize: plan the stages only, never write the vault
--limit <value>	Stage workers: cap the work processed this pass (segment, extract, synthesize-stage, link)
--convs <value>	Stage workers: comma-separated conversation ids to restrict the pass to
--at <value>	schedule: local wall-clock time for the nightly job, 24h HH:MM [default: 03:00]
--host <value>	Run the model calls on a serving mlx-bun at this host instead of loading the memory task model (127.0.0.1 when only —port is given)
--port <value>	Port of that server (8080 when only —host is given)
-h, --help	Show help

Use mlx-bun setup --help for terminal help.