inksprite v0.1.0-dev.13

Claude, OpenAI and local servers built in, narration through them, and
three guides.

• Anthropic, OpenAI, LM Studio, Ollama, llama.cpp, vLLM and KoboldCpp
  are built in beside OpenRouter, each local server at the address it
  listens on unless told otherwise, so connecting one is choosing it
  and a model. A hosted service needs only its key. Claude is spoken to
  through Anthropic's own API, its thinking shown and handed back on
  later rounds, and its prompts cached.
• The generation settings show only what the connection and model
  take, and only those are sent; a stored choice a model can't take is
  sent as the nearest it can. The reasoning effort menu gains Minimal,
  Extra high and Max for the models that have them. Max tokens and
  Tool rounds say Unlimited at 0.
• A server is asked for its models on Connect and Refresh, not as its
  address is typed. A chat on a service with no key says where to add
  it.
• Narration speaks through the AI connections: Kokoro is built in, and
  OpenRouter and OpenAI speak too. A project picks its connection and
  speech model at the top of the narration panel's Voices, and
  Settings › Narration is gone. In the desktop app, a project kept in
  a folder keeps its audio there.
• The narration panel has a player of its own and lists its segments
  in a Segments section, with Generate, Generate a sample and
  Regenerate all as buttons. Filters take regular expressions for what
  to leave unread, such as page numbers and reference marks.
  Pronunciation has a Preview, and the panel shows whole with no
  document open. A current Kokoro-FastAPI's voices are listed again.
• A turn's buttons are in a row under it, on the right, with a reply's
  other answers beside Regenerate. Sending the writer's turn again
  keeps the reply as another answer.
• A chat's system prompt and author's note open in a tab of the
  editor, from their fields in the chat's settings.
• Interpret and Write are in an Experimental group, off on every
  built-in profile and on existing chats and profiles; switched on,
  the model can call Write. Roleplay profiles start with the document
  tools on, so a card's lorebook is looked up rather than carried
  whole. Two dice rolls in one round both roll.
• The welcome is laid out as the settings are, with the AI settings
  and the Terms of Service, which are agreed to beside Start writing.
  About has the copyright, and the version on every build.
• The Jobs button shows only while there are jobs, and an empty
  folder no longer says "Empty".
• The docs site has guides to getting started, narration, and
  roleplay with a character card.

Downloads

A development build of the desktop app.

• macOS: the .dmg ending in arm64 for Apple silicon, or in x64 for an Intel Mac. Signed and notarized.
• Windows: the .exe. It isn't signed, so Windows warns that it doesn't recognize the app: choose More info, then Run anyway.
• Linux: the .AppImage, or the .deb for Debian and Ubuntu. If the AppImage won't start on Ubuntu 24.04 or later, install the .deb, or start the AppImage with --no-sandbox.

Once installed, the app updates itself: it downloads each new release and offers to restart into it. The .zip, .blockmap and .yml files are for the updater.

An app installed from a release before v0.1.0-dev.8, the first built with Electron, doesn't update to this one; install it over that one.