Velora 0.4.0

New

• Always on. Velora keeps the Mac awake while it serves (and takes fewer requests at once when the Mac is hot or in Low Power
  Mode), runs headless with Velora --serve, can open at login, and the menu bar shows whether it is idle, serving or throttled.

• Reach it from your phone. The gateway can listen on your network (HTTP by default, HTTPS with a self-signed certificate as an
  option), advertises itself on Bonjour, and shows a QR code with the URL and a key.

• Search and ask your files. Add a folder (PDF, text, Markdown, HTML, RTF, source) under Knowledge: it is indexed on your Mac
  with a local embedding model, searched by meaning, and used in the Playground or through the gateway to answer with sources.
  /v1/embeddings is available to your own tools.

• Quick Ask. Control-Option-Space opens a floating prompt in any app; select text and use Services > Ask Velora about Selection.

• Shortcuts, Siri and Spotlight. Ask Velora, Search My Files, Start Gateway and Stop Gateway.

• Serving that behaves. Memory is reserved for the KV cache and prompt cache from each endpoint's context length and parallelism;
  extra requests wait in a bounded queue; idle models unload and reload on demand; token usage per day, endpoint and API key; a
  Prometheus /metrics endpoint; Benchmark and Test Tools (tool-calling check) buttons.

• Air-gapped Macs. Export models and the MLX runtime to a checksummed folder and import them on a Mac with no internet.

• Kubernetes on slow links. The k3s binary is downloaded once on your Mac (eight parallel connections, SHA-256 checked) and handed to the cluster VMs, instead of each VM fetching 81 MB from GitHub.

• Home, snapshots and clones. A Home screen (memory, machines, endpoints, gateway, next steps); instant copy-on-write snapshots and clones of stopped machines (APFS), with restore.

• Model browser. Search MLX models on Hugging Face (or paste a repo) under Models; Velora reads each model's config.json and file sizes to work out its memory and KV cache and whether it fits, before you download.

• Anthropic and Ollama APIs. The gateway answers /v1/messages (Claude Code, Anthropic SDKs) and /api/chat, /api/generate, /api/tags (Open WebUI and other Ollama clients), streaming and tool calls included.

• Doctor. A Host screen that checks disk, memory pressure, heat and Low Power Mode, the runtime, downloaded models and endpoints, and times the download hosts (Hugging Face, GitHub, Debian, Ubuntu, Apple) so a slow route is named before a download stalls.

• Shared folders. Share a Mac folder (optionally read-only) with a machine over virtio-fs; Linux guests mount it by tag, macOS guests see it under My Shared Files. The machine view now has Overview, Snapshots and Sharing tabs.

• Rosetta for Linux. A switch in the machine's Sharing tab runs x86_64 programs and containers in an ARM64 Linux guest; Velora sets it up inside the guest over SSH and keeps it across reboots.

• Suspend to disk. Suspend saves a running machine's memory to disk and frees it; starting it again resumes exactly where it was (a program's state and RAM-only files survive). Linux guests now keep one machine identifier between starts, which saved state requires.

• Edit and resize. Rename a stopped machine and change its CPUs and memory; grow its disk (Linux guests from Velora's images use the new space at the next start).

• Port forwarding and bridged networking. A new Network tab forwards a port on this Mac (localhost only) to a port in a NAT guest, live while the machine runs. Machines can also be bridged to a physical interface to get their own LAN address (untested here: needs Apple's networking entitlement and a LAN).

• Clipboard sharing. Linux machines can share copy and paste with this Mac through a SPICE agent channel; Velora installs the guest agent over SSH.

• Control API and velora command. /v1/admin/vms starts, stops, suspends, snapshots, restores and runs commands in machines, behind a token only you can read. Settings installs a velora shell command (ls, start, stop, exec, ssh, snapshot, restore) that uses it.

• Sandbox VMs. velora sandbox run -- <command> (and POST /v1/admin/sandbox/run) clones a prepared Linux base, boots it in about 5 seconds, runs one command with CPU, memory and time limits, and deletes it. Linux only.

• MCP server. Velora --mcp lets Claude Desktop, Claude Code and other agents list and start machines, run code in a sandbox, search knowledge collections and ask local models. Settings has a Copy MCP Config button.

• JSON mode. response_format with json_object or json_schema on chat completions: the answer is checked against the schema and retried once with the error; a miss is a clear 422.

• Vision and speech add-on. A separate mlx-vlm environment (installed on first use, so the text runtime stays small) serves image models and speech to text: images in chat messages, and /v1/audio/transcriptions. Catalog adds SmolVLM, Qwen2-VL, Whisper Tiny and Small, and Soprano (text to speech at /v1/audio/speech; translations too). The Playground can attach an image, and the offline runtime bundle carries the add-on.

• Model aliases, fallback, A/B and speculative decoding. An endpoint can answer to other names (swap the model behind a name an app already uses), fall back to another endpoint when it cannot answer, send a share of traffic to another endpoint (callers stay on one side), and (experimental, slower in our test) use a small draft model for speculative decoding. Every response says which endpoint served it in X-Velora-Endpoint.

• Keys, budgets, charts and an audit log. API keys can expire, have daily token and request budgets, and be edited (name, endpoints, rate, expiry, budgets) without changing the secret. A new Usage screen (Swift Charts) shows tokens per day by endpoint, decode speed over recent requests, and a filterable, exportable (CSV) log of every gateway request: who, which route, status, time, tokens; never the prompt, the answer or the key.

• Saved chats and prompts. Playground and Quick Ask conversations are saved on this Mac (one file each, readable only by you) and listed under Chats: search, rename, delete, export or copy as Markdown. The Playground has History, New Chat and Export; a Prompts menu inserts reusable prompts (a few starters; add, edit and delete your own). Settings has a switch to stop saving and a Clear button.

• Welcome and a setup checklist. On first launch Velora asks what you want it for (virtual machines, local AI, agents and automation). Home then shows a checklist for those goals that ticks itself from what is true on your Mac (a machine exists, the runtime is installed, a model is downloaded, the gateway is running…), with a Go button for the next step, Skip for any step, and Hide. Help > Welcome to Velora and Settings reopen it. Nothing is installed or downloaded until you press Go.

• Export and import machines. Overview > Move or Back Up > Export packs a stopped machine into one folder (<name>.vmbundle: the disk, firmware and identity files with a checksum for each); File > Import Machine… (or dropping the folder on the window, or velora export and velora import) brings it back as a new machine with its own identity and network address. Checksums are verified before anything is copied; a bundle that lists files it should not, links, or paths outside itself is refused. The disk stays sparse (a 32 GB disk with 1.2 GB of data exports as 1.2 GB), and a bundle copied through a file system that is not sparse still verifies.

• Console & Command. A running Linux machine's toolbar opens its serial console (live), a command box and file upload and download. Commands and files go through a small guest agent over virtio-vsock, so they work with the guest's network down; guests from older seeds fall back to SSH, and Install Agent puts the agent into them over SSH.

• FluxVM, wired further. New VMs take a named image (Debian, Ubuntu, Fedora, CentOS Stream, AlmaLinux, Rocky), port forwards and shared folders. Running VMs get the same console, command and file sheet through FluxVM's API, snapshots (take, restore, delete), and stopped VMs can be cloned (APFS copy, new name and MAC address).

• Kairon. A new Infrastructure screen registers this Mac as a Kubernetes Node in a Velora k3s cluster; Machines scheduled to it boot through FluxVM.

• Fleet cards. Each Mac in the fleet, this one included, is a card with memory and memory pressure, heat (macOS's thermal state), disk, Thunderbolt links and RDMA status, and per-endpoint speed and request counts, refreshed every 5 seconds.

• EXO (experimental). Fleet can start EXO on every node in one namespace and place a model across the Macs; the endpoint version is served by EXO behind the gateway's keys, rate limits and versions, with rollback. Verified for process wiring on one Mac only.

• Notarized releases and an installer package. Releases ship a .pkg installer next to the DMG (scripts/make-pkg.sh). Tags v* build both in CI, signed and notarized when Developer ID secrets are set, and publish it to zyvorai/velora; check.sh fails a built app that lost its virtualization entitlement.

Licence and macOS

• Velora is now under the Business Source License 1.1: free for personal and home use, learning, research, labs, development
  and testing; production business use needs a commercial licence (see LICENSING.md). Builds before 0.4 (0.3.x) were Apache-2.0.
• macOS is licensed by Apple, not by Velora (MACOS-LICENSING.md). The app limits you to two active macOS guests by default and
  asks you to confirm Apple's terms the first time you start one.

Fixed

• Output of programs Velora runs could...