Calibre 9.15 arrived on 18 September 2026 with an unusual headline feature: an optional "Create your own Adventure" writing game that uses an AI backend to manage a fictional world while the user drives the narrative. The feature began as a testbed for Calibre's AI integrations but grew into a playable choose-your-own-adventure experience, according to developer Kovid Goyal.
The game is completely optional and is not loaded unless the user explicitly enables it. To access it, open Preferences → Toolbars & menus and add "Create your own Adventure" to the main toolbar. Once launched, users can choose from eleven preset settings — including fantasy, space opera, cyberpunk, western, and gothic horror — or write a custom world description. The AI then generates an opening scene and a cast of characters; the player types actions in free text, and the AI narrates the consequences turn by turn. A "Read" button exports the ongoing story to an EPUB file with a cover, tags, and chapters ready for any e-reader.
Confirmed
- Release date: 18 September 2026 (version 9.15).
- New feature: "Create your own Adventure" interactive storytelling game — optional, off by default, and not loaded until explicitly run from the toolbar.
- Eleven preset world templates plus free-form custom world input; story export to EPUB with cover, tags, and chapters.
- Nine AI provider options spanning cloud APIs (including Anthropic, Google, Grok, and OpenRouter) and local runtimes (Ollama and LM Studio).
- Viewer: multi-select highlights with style changes; optional sidebar button for saved window sizes; highlight notes as hover tooltips (respecting the existing tooltip preference).
- Read Aloud: no longer stops at some full-screen images; better text visibility in flow mode.
- EPUB 3 upgrades keep landmarks; DOCX output preserves hyperlinks across styled runs; DOCX import fixed for numbering without styles; spell-check handles typographic single apostrophes.
- Edit Book: snippet dialog test-box error fixed; plugin toolbar icons can be moved or deleted again; calibredb embed_metadata no longer skips books without covers.
- Cover grid: choose which corner holds cover emblems; Windows cover-grid background-color changes now apply.
- Windows startup fixed for Alberta, Canada (new timezone); Content Server sanitizes supplied conversion options; OpenRouter image-only models are no longer listed.
- New CLI flag can add or delete a book from a specified path in the system temp directory when launching calibre.exe.
- New news source: Briefing Service by Jordan Shelley; improvements to The Week, Bloomberg, and Global Times.
Unknown
- There are no independent benchmarks for AI response latency or token costs across the nine providers. One local hands-on test with Ollama and a ~23 GB Qwen model on a Mac M4 reported roughly one minute per turn for about 450 words of prose; results will vary by hardware and model size.
- Prompts sent to the model are deliberately not translated. World descriptions written in languages other than English may therefore get English replies unless the user explicitly asks for another language during setup.
- There are no SLA, pricing, or privacy guarantees for the cloud AI backends. Local backends such as Ollama and LM Studio keep data on-device but need enough GPU or CPU resources and RAM.
- Linux distribution repositories typically lag several versions; the project points users at the official binary installer.
Our take
Calibre's AI storytelling game is a clever dogfooding exercise: it turns the project's AI abstraction layer into a user-facing feature without forcing it on users. The opt-in design and local-provider support from Ollama and LM Studio fit privacy-conscious Calibre readers. The real test is whether that abstraction stays stable as upstream model APIs shift — a pressure the project already felt when it dropped the discontinued GitHub backend in 9.14.
