Messages stream in real time, token by token, with a bespoke quill-writing animation that gives way to rendered Markdown as the response completes. A status indicator above the composer shows the current processing stage—compressing context, gathering memories, building the prompt, streaming the response, executing tools—so you always know what the system is doing and why it has not answered yet.

The composer is keyboard-friendly, auto-resizing, with an inline Markdown preview toggle and paste-to-attach for images. Attachments appear in the footer with per-provider file type awareness. If a provider does not support native file attachments, a cheap-LLM fallback generates text descriptions for images and inlines text files, streaming status events so you know what is happening. Draft messages persist automatically—close the tab, come back tomorrow, and your half-written thought is still there.

Server-rendered Markdown handles the heavy lifting for simple messages, with client-side rendering as a fallback for messages with embedded tools or attachments. Roleplay bracket patterns render distinctly from dialogue. Code blocks get syntax highlighting. Emphasis survives streaming. And messages starting with a tab character no longer render as preformatted code blocks, which was a bug that persisted longer than anyone would like to admit.

LaTeX math now typesets in the bubble, rendered with KaTeX—wrap it in double dollars ($$…$$) or the\(…\) and \[…\] forms models tend to reach for. A single dollar sign in front of a number is left strictly alone, so “$50 across the table” stays fifty dollars rather than being mangled into an equation. Display equations too wide for the message scroll sideways inside it rather than stretching the room out of shape.

Reasoning models—DeepSeek, Anthropic’s extended thinking, Gemini, OpenAI and Grok reasoning summaries—now show their chain-of-thought inline in the assistant bubble, streamed live as it happens: collapsible, offset to the right, rendered in dimmed italic so the thinking reads as marginalia rather than dialogue. Visibility is controllable per-chat and globally (show, hide, or start collapsed). Tool calls initiated by characters now splice into the prose at the exact point they fired, rather than stacking at the bottom of the bubble like a postscript nobody asked for—position is captured server-side and persisted across reloads. Staff announcements from the Host, Prospero, the Lantern, Aurora, and their colleagues render as compact importance-coded chips that flex-wrap onto a row: red dot for high priority, amber for medium, grey for low—no more parade of full-width banners. And single newlines now render as actual line breaks, matching the convention of every chat application built since roughly 2005 and fixing the long-standing issue where multi-line blockquotes collapsed into one breathless run-on sentence.

Auto-scroll on response completion is now opt-in, with the default set to off—the view stays exactly where you left it, which is where you were reading, which is where you want to be. A floating jump-to-bottom button appears when you have scrolled up, in case you change your mind. Sending your own message still scrolls to the bottom, because at that point you clearly want to see what happens next.

Starting a fresh conversation—or continuing one elsewhere—quietly does a great deal of slow work before the first line appears: resolving the cast, running a per-character “choose what to wear” step, compiling identity stacks, backfilling continuation history, and seeding the opening scene. The wardrobe step is usually the longest of them, one deliberation per character. Until now the app simply sat there while it happened.

Now a blocking, non-dismissable status dialog—the Green Room—appears the moment creation begins. It shows a live status line, and for each character choosing an outfit it opens a “consulting the wardrobe for Name” panel that resolves into the decided four-slot outfit—top, bottom, footwear, accessories—with a scrolling activity log running beneath. The dialog cannot be dismissed while creation runs; it closes itself the instant the conversation is ready for input, and offers a Close button only if something goes wrong.

The Salon was built from the ground up for multi-character interaction. A participant sidebar shows every character in the conversation, sorted by predicted turn order with numbered position badges: green pulsing for the character currently generating, green for next, blue for queued, amber for your turn. Each card offers a connection profile dropdown—now on every seat, including the one you type as, so you can hand any chair to an LLM or reclaim it for yourself—an active/inactive toggle, and expandable settings for system prompt overrides. Nudge a character to speak and the invitation is now a permanent part of the transcript: the Host turns to them and invites them to take the floor, posted as a real announcement rather than a note that vanished on reload.

A character who has just consulted its memory and looked something up should not then contradict either. Before a character’s tool-using reply is saved, an optional cheap-LLM check compares it against what the character was told this turn—its last Commonplace Book whisper—and what it looked up. A consistent reply is vouched and saved as it stands. An inconsistent one is shown the discrepancies and asked to either stand by its answer or rewrite it; the first reply streams live and is quietly replaced in place if the character revises it, a visible and deliberate act of transparency.

Each checked message carries a small badge—Vouched,Amended, Stood by, orUnvetted—that reveals the check’s notes on hover. The gate is off by default, with per-project and per-chat overrides: a chat’s own switch lives in the Salon sidebar, and the project-wide control sits with Prospero. When there is nothing to check, the check does not run, and nothing is written.

The Lantern generates AI-powered story background images that appear behind your chat content at 45% opacity, creating a sense of place for each conversation. But the experience of those backgrounds—the way they transform the Salon from a chat window into alocation—belongs here, because atmosphere is a property of the room, not the projector.

When a chat reaches a natural scene-setting moment, the system derives a scene context from recent messages—not a literal transcript, but an imaginative scene description. Characters are depicted as they currently appear, wearing what the narrative says they are wearing, in a setting that reflects the mood of the actual conversation. The result pins to the top of the viewport behind the chat content, visible as a thumbnail in the chat header and expandable to full-screen on click. Story background thumbnails appear on chat cards throughout the application, giving each conversation a visual identity before you open it.

Theme background images yield to story backgrounds automatically. Manual regeneration is available from the tool palette. For chats flagged by the Concierge, image generation reroutes to your configured uncensored provider—the atmosphere adjusts to the conversation, not the other way around.

A toolbar above the composer provides access to everything you might need during a conversation without leaving the chat. The palette is organized into labeled sections—Chat, Organize, Edit Content, Memory—with composer gutter tools for the actions you use most: attach files, generate images, and roll dice.

Roleplay templates govern the structural conventions of a conversation—how actions are denoted, how dialogue is formatted, what the LLM should and should not include in its responses. Templates are provided by plugins, making them shareable and versionable. Quilltap ships with a default template, and the plugin SDK includes a createRoleplayTemplatePlugin()builder for creating your own.

System prompts are also plugin-provided. Ten built-in prompts cover the major providers—Claude, GPT-4o, GPT-5, DeepSeek, Mistral Large—in both companion and romantic variants, each architecturally tailored to the provider’s strengths and failure modes. Characters can carry multiple named system prompts with a per-chat selector, so the same character can use different prompting strategies for different conversations.

Template variables—{{char}},{{user}},{{timestamp}}—are resolved at prompt assembly time. The template display highlights these variables and warns when hard-coded names appear where variables should be. Timestamp injection supports friendly, ISO, date-only, time-only, custom, and fictional time formats with per-character and per-chat configuration.

Conversations export in Quilltap’s native .qtapformat with selective entity inclusion and memory options, or in SillyTavern-compatible JSONL for interoperability. Import supports both formats with conflict detection and three resolution strategies: skip, overwrite, or duplicate. Post-import reconciliation updates all foreign key relationships automatically.

SillyTavern import handles multi-character conversations with a speaker mapping wizard that lets you assign every speaker—user and AI alike—to any available character. Newly imported chats receive a highlight animation in the sidebar so you can find them immediately.

The full backup system captures everything—characters, chats, memories, files, plugin configurations, and npm-installed plugins—in a single ZIP archive. Restore recreates your entire installation from that archive, with entity remapping for fresh-account scenarios.

Provider errors do not produce cryptic failures. When a request exceeds an LLM’s limits—too many tokens, too many PDF pages, an image too large—the system attempts graceful recovery: a simplified message explaining what happened is sent to the LLM, which provides an in-character response suggesting alternatives. A two-tier fallback ensures the user always sees something: LLM-generated recovery first, then a static fallback if the recovery itself fails.

Streaming errors no longer cause messages to vanish from the UI. If a provider returns an error mid-stream, the user message is preserved (it was already saved server-side) and the chat re-syncs to reflect the actual state. Keep-alive pings during long operations like context compression prevent proxy and load balancer timeouts.

When a provider silently refuses content—returning an empty response instead of an error—the Concierge catches the silence and retries with the same provider first (in case of a transient issue), then fails over to an uncensored provider if the silence persists. Greeting generation, memory extraction, story backgrounds, and every other background task receive the same treatment. The Salon does not go quiet without a fight.

An Enclave is an all-AI conversation that runs on its own schedule or on demand, without the operator sending messages. Set up a room, populate it with characters, define the premise, and step back. The characters talk to each other—planning, arguing, telling stories, solving problems—while you watch, or while you are away doing something else entirely. You can read the transcript later, like finding a stack of letters someone left on your desk.

Budget controls keep things from running away with your token allocation or your credit card. Set caps on turns, tokens, wall-clock time, and estimated spend—any combination, any threshold. Cache-read tokens can optionally be excluded from budgets, so efficient caching is not penalized. The Host posts pacing announcements at the halfway and near-end marks so characters have time to wrap up gracefully, and a grace turn is granted when the budget is hit without prior warning, because even fictional people deserve to finish their sentences.

Release 4.7 wired the corridors between the rooms—thePost Officefor letters between characters, theBrahma Consolefor the keeper—and along the way it polished a great many small edges in the Salon itself.