Pasta-Devs/Marinara-Engine v1.5.6 on GitHub

Added

New connection provider Claude (Subscription) that routes chat through the locally installed Claude Agent SDK so requests bill against your Anthropic Pro / Max subscription instead of an sk-ant-* API key. Requires npm i -g @anthropic-ai/claude-code and a one-time claude login on the host running Marinara. This is the same auth mechanism Anthropic-endorsed integrations like Zed use; no proxy or third-party shim is involved. Built-in agent tools are disabled and use Marinara's own agent/tool layer. Embeddings are not supported on this provider; configure a separate connection for them.
The "Mari is thinking…" indicator appears above the composer while Professor Mari executes her embedded commands (create/update character, fetch, create chat, navigate). Makes it clear that her background work is running, not frozen. Bonus: Dottore is doing jumping jacks.
Dry-run generation endpoint (POST /api/generate/dryRun) that runs the full generation pipeline without side effects; no messages persisted, no agents or tools invoked, no Discord webhooks. Extensions can send a userMessage to preview "what if I said this", use impersonate: true to preview the user's next in-character line, enable optional injections (lorebook, trackers, chat summary), override the preset or connection, and optionally receive the assembled prompt instead of a completion (returnPrompt: true). Supports both non-streaming JSON responses and SSE streaming with abort capability. Intended as a stopgap extension API for flexible prompt inspection and silent generation.
In Game mode, NPCs can be added/removed from your party, plus now you can manage the party manually.
If you have Image Generation enabled in Game mode, during important scenes, the model now generates immersive VN-like scenes from the player's POV.
Overall improvements to generating expressions/full-body sprites for your characters.
Guided generations with a visible indicator.
Schedule generation preferences added for conversations.
Pygmalion, Jenny, and DataCat added to the Browser.
Pinnable taskbar shortcut via custom launcher.
Universal Tool Support for agents.
New Knowledge Router agent.
You can now link Personas to Lorebooks.
Drag-and-drop Lorebook entries.
Added ElevenLabs for TTS support.
You can now see spoilers for Game mode and edit the plot accordingly to your needs in the History section.
Upon ending the Game session, you can now optionally include what you want to happen in the next session.
Separate volume levels for different sounds in Game mode.
Added the /impersonate_prompt command that allows you to change the impersonate prompt.
Added manual mode in Conversations that only makes the character respond when you ping them with @name.
Resizing sprites in game mode.

Fixed

UI and other minor glitches in Game Mode.
Image Generation in game mode is not firing up for named NPCs in a scene.
More ComfyUI fixes.
Various general fixes and improvements.
Anchor link error.
We now enable the send button immediately after branching.
Remove background actually sticks across switches.
Sidecar CUDA runtime setup fix.
Light Mode readability issues.
Removed the ability to apply presets to Conversations, which broke the format.
Improved usability on mobile devices with small screens, where tapping tiny buttons could be difficult.
Navigational icons under messages now scale with the display size.
When selecting Personas during chat setups, you can now see their avatars.
Switching between chats doesn't cancel generations in progress.
Parameters added to Conversations and Roleplay setups.
Bugged NPC entries in Game mode journal.
Creating a new agent doesn't delete the old one.
Preset names are no longer set to Default upon being selected.
Black screen on search bar typing in chats was fixed.
Various UI fixes applied.
DeepSeek V4 is now supported.
Addressed the bug that deleted your Persona fields when uploading an avatar in an unsaved state.
Minor adjustments to some agent widgets.
Game mode now supports multiple maps.
Debug mode restored.