mere.run
Local multi-modal AI inference runtime and CLI
TLDR
SYNOPSIS
mere.run [global-options] command [subcommand] [options] [args...]
DESCRIPTION
mere.run is a local-first inference runtime and public CLI for multi-modal generation and understanding. One command covers image, text, speech, vision, music, sound, video, 3D-related reconstruction helpers, model management, and an optional OpenAI-compatible API server. The optional macOS Studio app uses the same CLI, model store, and run history rather than a separate backend.The CLI is modality-first: creative and analysis work lives under groups such as image, text, speech, vision, music, sfx, and video, while model, adapter, api, status, and plugin handle operations. Managed model IDs are pulled into a local store (by default under Application Support on macOS); use model capabilities before large downloads. Packaged offline cookbooks are available via mere.run guide.Primary development targets Apple Silicon macOS; headless Linux CLI packages are supported for the runtime without the macOS SwiftUI studio. Media workflows typically expect ffmpeg and ffprobe on PATH (overridable with MERERUN_FFMPEG / MERERUN_FFPROBE).
PARAMETERS
setup
Guided first-run setup for the local runtime and model store.guide [command-path]
Read packaged offline cookbooks for a command path; use guide --list to list them.model {list, pull, remove, info, capabilities, runtime, benchmark, ...}
Manage the local model store: pull by managed ID, inspect install status, and report what the machine can run.image {generate, train-lora, run-plan, validate, ...}
Local image generation, LoRA training, and related image workflows.text {chat, code, embed, anonymize, train-lora}
Local chat, code generation (llama.cpp), embeddings, and PII redaction.speech {synthesize, transcribe, diarize, listen, profile}
Text-to-speech, ASR, diarization, and voice profiles.vision {caption, inspect, face, ground, segment, track, pose, ocr, ...}
Captioning, VQA, face analysis, grounding, segmentation, tracking, pose, and OCR.music {analyze, generate, realtime, transcribe}
Music generation, analysis, realtime performance, and MIDI transcription.sfx {generate, video, ae, clap, ...}
Sound-effect generation, including video-conditioned Foley.video {generate, animate, prepare-masks, session, ...}
Local video generation and subject animation pipelines.api serve
Serve OpenAI-compatible chat, embeddings, images, TTS, and STT on a local host/port.status
Show local server, served model, model-store path, and installed models.plugin {list, info, install, doctor}
Discover and manage companion plugin executables.--models-root path
Override the local model store directory for this invocation.--help
Show the public command tree or help for a group/subcommand.
CONFIGURATION
~/Library/Application Support/MereRun/models
Default local model store on macOS. Override with MERERUN_MODELS_DIR or --models-root.MERERUN_HUB_CACHE / MERERUN_MODEL_CACHE_HOME
Control the Hugging Face snapshot cache used by model pull and auto-download paths.MERERUN_API_KEY
API key for non-loopback api serve binds; loopback can run without auth by default.MERERUN_FFMPEG / MERERUN_FFPROBE
Absolute paths to ffmpeg/ffprobe when not discovered on PATH.Use mere.run config {get, set, unset, list, path} for runtime configuration next to the active model store.
CAVEATS
Large models need substantial RAM/unified memory; check model capabilities before pulling. Licensing, download size, and hardware fit vary by model ID—see project model-sources docs before redistributing weights. Linux support is headless CLI-oriented; the SwiftUI studio and DMG installer are macOS-only. Binding api serve beyond loopback requires an explicit API key. Tool-loop and shell features in chat require explicit opt-in flags that weaken default safety posture.
HISTORY
mere.run is an open-source Swift/MLX local inference runtime and CLI by sawfwair, packaging multi-modal creative workflows (image, text, speech, vision, audio, video) behind a single mere.run executable, with optional macOS Studio and Linux headless packages.