mosaicGetting StartedLaunch the interactive TUI coding agent in the current workspace directory.
mosaicA focused reference for autonomous coding commands, hardware acceleration, remote machines and everyday App workflows.
mosaicGetting StartedLaunch the interactive TUI coding agent in the current workspace directory.
mosaicmosaic .Getting StartedOpen the current directory with auto-detected local model configuration.
mosaic .mosaic chatGetting StartedStart a conversational chat session without workspace tools, shell execution, or web access.
mosaic chatmosaic "<instruction>"Getting StartedExecute a headless autonomous prompt directly from your terminal and exit upon verification.
mosaic "Fix memory leak in buffer pool and run test suite"mosaic machine linkGetting StartedLink the current machine to your account and automatically prepare its compatible Crea image runtime, or add --anon to share a restricted machine with Methil members without GitHub authentication.
mosaic machine link --anon--creative-model <id>Getting StartedDeclare a creative specialization for a linked machine. Repeat the flag to make one worker serve several catalogued models; omit it to expose the complete catalog.
mosaic machine link --anon --creative-model krea-2-turbo --creative-model ltx-2.5/setupSlash CommandsOpen the hardware auto-configuration wizard to benchmark GPU/RAM and pick optimal models.
/model <id>Slash CommandsSwitch the active language model on the fly (e.g. /model qwen3.8-27b-q4_k_m).
/provider <name>Slash CommandsSwitch provider route (local, lmstudio, openai, openrouter, anthropic, ollama).
/doctorSlash CommandsRun comprehensive workspace diagnostics, verify build tools, git tree, and runtime dependencies.
/statusSlash CommandsDisplay current token usage, active provider, context budget, and system resource telemetry.
/diffSlash CommandsView pending file changes and uncommitted git modifications generated by the agent.
/reviewSlash CommandsInitiate an automated code review pass on recent edits before committing.
/newSlash CommandsReset conversational context and start a fresh coding session.
mosaic machinesRemote Machine ControlList all linked compute nodes, online statuses, accelerators, and memory capacity.
mosaic machine listmosaic machine statusRemote Machine ControlInspect local machine registration, system daemon health, and telemetry heartbeat.
mosaic machine statusMosaic Crea model queueRemote Machine ControlFrom App / Machines, enqueue an allowlisted Krea 2, MiniMax H3, MiniMax Music3 or LTX-2.5 download on a compatible online worker. Progress and logs stay attached to the preparation job.
mosaic machine link --anon --creative-model minimax-music3mosaic machine rotate-tokenRemote Machine ControlRotate the cryptographic access token of this computer and re-authenticate.
mosaic machine rotate-token--provider <provider>CLI Flags & OptionsOverride default model provider (local, openrouter, openai, anthropic, deepseek).
mosaic --provider openrouter --model deepseek/deepseek-v4-pro--workspace <dir>CLI Flags & OptionsSet target root directory for workspace inspection and tool executions.
mosaic --workspace ~/projects/backend--jsonCLI Flags & OptionsFormat output as structured JSON / JSONL for CI/CD pipeline automation.
mosaic doctor --probe --json--no-creativeCLI Flags & OptionsDisable the creative model catalog for this machine when it should only provide language-model inference.
mosaic machine link --no-creative--no-imageCLI Flags & OptionsSkip automatic Crea image runtime preparation during machine link when the host should only provide language-model inference.
mosaic machine link --no-imageApple Silicon MLXHardware & RuntimesNative unified memory acceleration for M1/M2/M3/M4 Macs. Zero VRAM overhead, instant startup via HuggingFace MLX community.
llama.cpp (Metal & CUDA)Hardware & RuntimesUltra-portable GGUF inference supporting split CPU/GPU layer offload with quantized KV cache.
SGLang & vLLMHardware & RuntimesHigh-throughput server engines for Linux NVIDIA CUDA and AMD ROCm with FP8 / MXFP4 support.
Mosaic dynamically detects your system's memory topology at startup. On Apple Silicon, Darwin virtual memory counters (vm_stat) are parsed to distinguish active wired memory from reclaimable page caches, helping quantizations like Qwen3.8 27B or Qwen3.5 4B MLX fit safely in available RAM without kernel OOM panic.
On Linux workstations with NVIDIA or AMD hardware, Mosaic queries PyTorch and ROCm sysfs counters to balance layer offloading between dedicated VRAM and host RAM.