Hildegard 0.1.6 (pre-alpha)

Hildegard 0.1.6 is a pre-alpha. It brings Hildegard into your AI apps as the way they reach your own computers, with
its craft skills and a loop that keeps improving a piece until it passes.

Install

• macOS and Linux: curl -fsSL https://dbaney.com/hildegard/install.sh | sh
• Windows (PowerShell): irm https://dbaney.com/hildegard/install.ps1 | iex
• Desktop app: pick your system's download below, or from dbaney.com/hildegard.

Installed already? Hildegard updates itself, or run hildegard update.

New

• A Claude Code plugin. /plugin marketplace add baney75/hildegard, then /plugin install hildegard@hildegard:
  Hildegard's tools on your GPUs and every skill, in two commands. If Hildegard isn't installed yet, the plugin tells
  Claude how to install it.
• Every chatbot's steps in docs/chatbots.md:
  the Claude app, claude.ai and Cowork, ChatGPT, Grok and Grok Bot. Below are hildegard.plugin (skills for the
  Claude app and Cowork) and a skill-*.zip for each skill (claude.ai and ChatGPT uploads).
• Skills over MCP. Any app linked to Hildegard can read its skills with list_skills and read_skill, and apps
  that list MCP prompts show each skill as one.
• The hildegard skill: the midwife's method for serious creative work in any form, from a short interview and real
  research to drafts, contact sheets, blind critique and checks.
• 3D models. The hildegard-3d skill builds models in code (mesh_kit.py), and Hildegard checks each one:
  watertight, manifold, winding, volume, loose bodies, debris and size, and for prints units, bed, base, overhang and
  balance, with a sheet of eight views, a top or bottom view and silhouettes.
• The loop. loop_start, loop_submit, loop_critique and loop_finish take a 3D model, still, take or sound
  through generate, render, critique, refine. Hildegard takes the next round only after every question about the
  last one is answered with what the sheet shows, and keeps every round.
• A coding agent can run a whole loop for you: hildegard loop run codex "a heron finial, 120 mm, for my printer" --kind model --print. It runs within each provider's terms: started by you, on a plan in its own app, or on your
  API key.

Known limits

• Macs don't render on their own GPU yet. Pair a Mac with a computer that has an NVIDIA or AMD GPU
  (hildegard choir join).
• The Windows and macOS apps aren't code-signed yet, so Windows SmartScreen and macOS Gatekeeper ask you to confirm the first launch.
• claude.ai, ChatGPT, Grok and Grok Bot reach Hildegard only at a public link (hildegard host --funnel). Hildegard
  Link, the encrypted relay that needs no Tailscale, is coming soon.
• 3D models are checked from GLB, glTF, STL, OBJ and PLY with embedded data; Draco-compressed GLB isn't read yet.
• LTX models follow Lightricks' license: free under $10M yearly revenue, and not for products that compete with
  Lightricks. Wan 2.2 is Apache-2.0.

Hildegard 0.1.5 (pre-alpha)

Hildegard 0.1.5 is a pre-alpha. From this release on, Hildegard and its apps update themselves.

Install

• macOS and Linux: curl -fsSL https://dbaney.com/hildegard/install.sh | sh
• Windows (PowerShell): irm https://dbaney.com/hildegard/install.ps1 | iex
• Desktop app: pick your system's download below, or from dbaney.com/hildegard.

Installed already? Run the line above once more, or download the app again: this is the last time you'll need to.

New

• Hildegard updates itself. Every six hours it looks for a new release. When one is out and nothing is rendering or downloading, it installs it and starts again as the new version, keeping the one before. Choose when instead in Settings › About (Ask me first), or with hildegard update auto off; then hildegard update or Update in Settings when you want it.
• The desktop apps update themselves on macOS, Windows and Linux (the AppImage). A new app downloads in the background, and Restart to update appears at the top of the tray menu. The .deb doesn't update itself yet.
• Every update is signed. The apps and hildegard update refuse anything not signed with Hildegard's key, and refuse an older release passed off as a new one.
• The Omarchy panel offers Update to a new release, and says when Hildegard is updating.

Known limits

• Macs don't render on their own GPU yet. Pair a Mac with a computer that has an NVIDIA or AMD GPU
  (hildegard choir join).
• The Windows and macOS apps aren't code-signed yet, so Windows SmartScreen and macOS Gatekeeper ask you to confirm the first launch.
• Hildegard Link, the encrypted relay for reaching Hildegard without Tailscale, is a prototype and not switched on yet.
• LTX models follow Lightricks' license: free under $10M yearly revenue, and not for products that compete with
  Lightricks. Wan 2.2 is Apache-2.0.

Hildegard 0.1.4 (pre-alpha)

Hildegard 0.1.4 is a pre-alpha. It fixes a crash on start after upgrading to 0.1.3.

Install

• macOS and Linux: curl -fsSL https://dbaney.com/hildegard/install.sh | sh
• Windows (PowerShell): irm https://dbaney.com/hildegard/install.ps1 | iex
• Desktop app: pick your system's download below, or from dbaney.com/hildegard.

Fixed

• Starting after an upgrade. 0.1.3 stopped on start on a computer with a model installed but not the safety check, which is most upgrades from 0.1.1 and 0.1.2. Hildegard now starts first and fetches the safety check in the background, and a failed download shows on the model instead of stopping Hildegard.

Known limits

• Macs don't render on their own GPU yet. Pair a Mac with a computer that has an NVIDIA or AMD GPU
  (hildegard choir join).
• The Windows and macOS apps aren't code-signed yet, so Windows SmartScreen and macOS Gatekeeper ask you to confirm the first launch.
• Hildegard Link, the encrypted relay for reaching Hildegard without Tailscale, is a prototype and not switched on yet.
• LTX models follow Lightricks' license: free under $10M yearly revenue, and not for products that compete with
  Lightricks. Wan 2.2 is Apache-2.0.

Hildegard 0.1.3 (pre-alpha)

Hildegard 0.1.3 is a pre-alpha. It fixes upgrades from 0.1.1, reaches Hildegard from your other devices and from chatbots, and works with the local models you already run.

Install

• macOS and Linux: curl -fsSL https://dbaney.com/hildegard/install.sh | sh
• Windows (PowerShell): irm https://dbaney.com/hildegard/install.ps1 | iex
• Desktop app: pick your system's download below, or from dbaney.com/hildegard.

Fixed

• Upgrading from 0.1.1 no longer makes an installed model read as not installed, on the GPU computer and on every device paired with it. Hildegard now fetches the 343 MB safety check on its own when an installed model is missing it.

New

• Hildegard's tools from anywhere. AI apps on your other devices reach Hildegard over Tailscale with a key: hildegard host key "<name>", then hildegard connect <app> --remote <link>. Keys stay in the keychain, never in an app's config. claude.ai and ChatGPT can sign in to Hildegard with OAuth, and every call runs with that person's own permissions.
• Your local models. Hildegard finds Ollama, LM Studio, llama.cpp, vLLM, SGLang, MLX, Jan, KoboldCpp and the Local AI gateway running on this computer (hildegard llm). hildegard llm direct "<idea>" lets one of them direct a film.
• Stricter provider rules. Only the owner can run Claude on the owner's account or key, invited people included. claude.ai, ChatGPT and Copilot plan addresses and plan tokens are refused as model endpoints, since those plans may only be used in their own apps.
• LTX's terms in Hildegard now say what its license asks: label LTX videos as machine-generated wherever you share them.

Known limits

• Macs don't render on their own GPU yet. Pair a Mac with a computer that has an NVIDIA or AMD GPU
  (hildegard choir join).
• The Windows and macOS apps aren't code-signed yet, so Windows SmartScreen and macOS Gatekeeper ask you to confirm the first launch.
• Hildegard Link, the encrypted relay for reaching Hildegard without Tailscale, is a prototype and not switched on yet.
• LTX models follow Lightricks' license: free under $10M yearly revenue, and not for products that compete with
  Lightricks. Wan 2.2 is Apache-2.0.

Hildegard 0.1.2 (pre-alpha)

Hildegard 0.1.2 is a pre-alpha. It adds safety checks, one-step Hugging Face sign-in, coding agents that stay within each provider's terms, and Runpod GPUs. It also fixes what device testing of 0.1.1 found.

Install

• macOS and Linux: curl -fsSL https://dbaney.com/hildegard/install.sh | sh
• Windows (PowerShell): irm https://dbaney.com/hildegard/install.ps1 | iex
• Desktop app: pick your system's download below, or from dbaney.com/hildegard.

New

• No sexual content. Prompts, models, LoRAs and model files are checked before anything renders, and every frame of every take is checked after. Refused takes show as "Refused". See docs/safety.md.
• Hugging Face sign-in with no setup. Sign in from the browser, or with a device code over ssh. If this computer is already signed in with the hf command, Hildegard offers to reuse that token in one click. Tokens stay in the system keychain.
• Coding agents within their providers' terms. hildegard agents runs Claude Code, Codex and OMP with Hildegard's tools. A Claude or ChatGPT plan runs only in its own app, only when you start the run, and at a gentle pace. Automation and schedules use API keys. See docs/agent-terms.md.
• Skills and quality checks. hildegard skills install adds film, editing, photo, art, sound, setup and cloud skills for your coding agent. hildegard review <file> checks a take for black or frozen frames, flicker, clipping, exposure and loudness.
• Runpod GPUs. hildegard cloud runpod up rents a GPU with your own Runpod key. It joins your devices, and it stops itself when idle.
• App icons now match each system: the macOS squircle, the Windows plate, and a Studio favicon that follows light and dark mode.

Fixed

• A second install, or a second data folder, no longer stops the Hildegard that's already running or takes its port. Each home's server is its own, and a new one starts on the next free port and says which.
• Windows: upgrading while Hildegard runs from a copied app folder (exFAT or FAT32 drives) now stops it first, so the old folder can be replaced.
• Doctor: it no longer asks for a Hugging Face account while your takes render on a paired device.

Known limits

• Macs don't render on their own GPU yet. Pair a Mac with a computer that has an NVIDIA or AMD GPU
  (hildegard choir join).
• The Windows and macOS apps aren't code-signed yet, so Windows SmartScreen and macOS Gatekeeper ask you to confirm the first launch.
• Hugging Face Jobs and Colab GPUs aren't supported yet.
• LTX models follow Lightricks' license: free under $10M yearly revenue, and not for products that compete with
  Lightricks. Wan 2.2 is Apache-2.0.

Hildegard 0.1.1 (pre-alpha)

Hildegard 0.1.1 is a pre-alpha. It fixes what a full test of 0.1.0 found on Windows, macOS, Linux, WSL2 and Omarchy.

Install

• macOS and Linux: curl -fsSL https://dbaney.com/hildegard/install.sh | sh
• Windows (PowerShell): irm https://dbaney.com/hildegard/install.ps1 | iex
• Desktop app: pick your system's download below, or from dbaney.com/hildegard.

Fixed

• Windows:
  • The installers check free space before downloading anything. When there isn't enough room, they print the exact command to install on another drive.
  • Installing to an exFAT or FAT32 drive now works. Before, unpacking, the app link and uv's Python each failed on such a drive.
  • uv's download cache stays in Hildegard's folder instead of filling C:.
  • hildegard start over ssh no longer stops the server when the session closes.
• Every system: the server stays up when the terminal that started it closes, and its log says why it stopped.
• macOS: the disk image opens without a license prompt, and the app's signature is valid. The app isn't notarized yet: the first time you open it, choose Open Anyway in System Settings › Privacy & Security.
• Desktop app: it now uses an existing install wherever its hildegard command points, instead of offering to install a second copy.
• Doctor on a Mac: it now points you to pairing a GPU computer, and stops asking for a Hugging Face account a Mac can't use.
• Paired renders: a take's render time no longer counts its wait in the GPU computer's queue.

Known limits

• Macs don't render on their own GPU yet. Pair a Mac with a computer that has an NVIDIA or AMD GPU
  (hildegard choir join).
• The Windows and macOS apps aren't code-signed yet, so Windows SmartScreen and macOS Gatekeeper ask you to confirm the first launch.
• LTX models follow Lightricks' license: free under $10M yearly revenue, and not for products that compete with
  Lightricks. Wan 2.2 is Apache-2.0.

Hildegard 0.1.0 (pre-alpha)

Pre-alpha. The first public build of Hildegard: your own film studio, on your own GPU. Expect rough edges, and
please report them.

Install

• The app: Hildegard_x64-setup.exe (Windows, no admin rights), Hildegard_aarch64.dmg (Mac with Apple
  silicon), Hildegard_amd64.AppImage or .deb (Linux). The app isn't signed yet, so Windows SmartScreen and
  macOS Gatekeeper will ask you to confirm.
• One line in a terminal: curl -fsSL https://dbaney.com/hildegard/install.sh | sh on macOS and Linux, or
  irm https://dbaney.com/hildegard/install.ps1 | iex in PowerShell. Then hildegard setup.
• Every file's SHA-256 is in SHA256SUMS, and the installers check each download against it.

What works

• Direct shots with camera language (bodies, lenses, apertures, moves, looks) on LTX-2.5, LTX-2 and Wan 2.2, through
  ComfyUI, on NVIDIA, or AMD on Linux.
• The Choir: a Mac or laptop renders on another of your computers' GPUs over Tailscale.
• Host mode: your studio at an HTTPS link on your tailnet, with owners and directors.
• An MCP server, so Claude, Codex, Cursor and other AI apps can direct shots on your GPU.
• Hugging Face linking with the token kept in your OS keychain.
• The Omarchy bar widget.

Known limits

• Macs don't render on their own GPU yet: on Apple's GPU the models come out blurred and take over an hour. Pair a
  Mac with a computer that has an NVIDIA or AMD GPU.
• LTX models follow Lightricks' license: free under $10M yearly revenue, and not for products that compete with
  Lightricks. Wan 2.2 is Apache-2.0.