A SillyTavern extension for metaphysics roleplay: possession, body swaps, mind control, hive minds, dolls, time stops and more.
It keeps the rules, the bodies and the prose straight, so the model stops getting them wrong.

It pairs with the Worldhopper presets for Claude and Gemini, and works with any other preset, Chat Completion or Text Completion.

🧩 Only what you're into. Every mode is opt-in and picked per character, and nothing is on until you pick it. If mind control is your thing, pick Mind Control and that's all you'll ever get: no possession rules, no body swaps, no boxes or cards for modes you didn't choose. Mix as many or as few as you like, and change them any time.

⚡ Keep it fast and cheap. Give the Engine its own model, separate from the one you roleplay with. Its jobs are small background ones, and the Ledger runs after every single reply, so a local model (KoboldCpp, Ollama and the like) or a cheap, fast API model from the Flash, Haiku or mini tier does them well and costs next to nothing. Your big roleplay model would work too, but every background call would cost big-model prices, and your next reply waits for the Ledger to finish.

• ✨ What it does
• 🚀 Install
• 🧭 Quick start
• 🎭 Modes
• 📒 Body Ledger
• 👥 Running several bodies
• 💥 Moment boxes
• ✍️ Editor
• 🪪 Card tools
• ⚙️ Every setting
• ⌨️ Slash commands
• 🧰 The regex scripts
• 🔒 Privacy
• ❓ Troubleshooting

1. In SillyTavern, open Extensions → Install extension, and paste:

   https://github.com/Coz-rp/SillyTavern-Worldhopper
   

2. Open Extensions → Worldhopper Engine → Settings → Models and pick a connection profile for each job: a local or cheap, fast model, not your roleplay model (see Models).
3. Open a character and pick its modes. That's it.

You need SillyTavern with the built-in Connection Manager (tested on 1.18), and a connection profile for the Engine's jobs (plug menu → Connection Profile), ideally pointing at a local or budget model. No special preset is needed: the Engine writes its own prompts, and a profile's preset only lends its sampler settings.

On first start the extension installs its lorebook, WH Metaphysics, and a handful of display-only regex scripts (see The regex scripts). Both update themselves with the extension. If you edit your copy of either, your edits stay, and the extension asks before replacing anything.

By hand: copy this folder into SillyTavern/public/scripts/extensions/third-party/ and restart SillyTavern.

1. Open a chat with a character.
2. In Extensions → Worldhopper Engine, tap ✏️ under Modes and pick what this character uses, or tap 🪄 to have some suggested from the card.
3. Play. After each reply the Ledger updates, and the strip above the chat shows who is where.

Every section of the panel has a ❔ button that explains it in a few lines, and every switch in Settings says what it does right under its name.

A mode is one kind of metaphysics, with its own rules: how it's done, what the subject experiences, who writes what. The rules live in the WH Metaphysics lorebook, the Codex. Its entries stay off everywhere, and Worldhopper switches on only the ones for the modes you picked, only in that character's chats.

• Picked per character (or per group chat). Tap ✏️ to pick, 🪄 to get suggestions from the card, and tap a picked mode to read what it means.
• Pick the fewest that carry the card. Every picked mode is written into every reply, so a mode the card only hints at does more harm than good.
• Modifiers (↳) build on another mode and bring it in with them: Multipossession brings in Possession.
• What gets added: the core rules for your modes; a one-line reminder per mode near the end of the prompt, where it carries the most weight; and detail entries that switch on when the chat touches their topic, like trance depth or a skinsuit going on.
• Your card wins. The card sets the specifics (how a power works, what it costs, what the subject remembers); a mode's defaults only fill in what the card leaves out.

After each reply, a model you choose reads the newest messages and updates a record of:

• 🧍 Bodies that aren't driven by their own mind: possessed, worn as a skin, absorbed into a hive, overwritten by a copy, puppeted, swapped. Also empty bodies, and minds with no body.
• 🧠 Minds a power has changed while they stay in their own body: hypnosis, mind control, perception rewrites, blank slates, drones, pets, dolls, limp bodies. Each with what changed, any triggers ("'bloom' drops her into a trance") and how they are right now.
• ⚡ Powers: who can do what, even when nobody is using it.
• 🏷️ Names: one person under several names (a code name, a host number).
• 🙈 Whether you know. Anything kept from you stays out of the narration, and out of the strip's details.

The record goes into the prompt just before the model writes, so names, pronouns and every change stay straight. When it's your doing, the model is told that only you set or change it; when it's done to you, the model plays your changed behaviour and leaves your inner experience to you.

The strip above the chat shows it at a glance: 🟢 teal when you're the one doing it (inside a body, or controlling a mind), 🟣 violet when someone else is, dashed when a body is empty or something is hidden from you. Tap it to edit.

"As …" badges on your messages show whose body you wrote them from.

Body History remembers past body changes: who was inside, what the owner remembers, what was kept or left behind. For a chat that started before the Ledger was on, Build history from earlier replies reads the replies that have no Ledger entry yet, once each. On a long chat it offers to read just the last 30 instead of the whole thing.

It never rereads the chat. After each reply the Ledger reads only the newest messages, alongside what it already knows.

Fixing mistakes: ✏️ opens the Ledger to edit; the next update builds on your fix. 🔄 rebuilds it from the card and the last few messages.

Which model: the Ledger runs after every reply, and your next reply waits for it, so give it your fastest, cheapest model: a local one, or the Flash, Haiku or mini tier. With only Timestop picked there's nothing to track, so it stays off.

With Multipossession, as a hive's original mind, or working puppets, you run more than one body at once. The body roster makes that easy:

• A card per body sits over the message box, in that body's colour, with what it's doing. Your own body is marked (you).
• Tap a card to start a line for that body, like [Jess] leans on the counter. The card for the line you're on lights up; cards that already have a line say so.
• A line with no name is you as a whole, and goes to whichever bodies it fits. Jess: at the start of a line works too.
• Your lines show in the chat as rows in the same colours.
• Your bodies are yours. The writer only ever does what you wrote with them, with nothing added. A body you gave nothing stays as you left it; if someone speaks to it, the reply stops and leaves the answer to you.
• Copies are their own people. With Propagation, copies of you are separate people the writer plays, so they never get cards.
• Past six bodies the cards shrink to chips, and the corner button shrinks them any time.

The Codex asks the model to mark each mode's big moment, and a display-only regex script turns the marked passage into a styled box. Only the moment gets a box (one to three paragraphs, never the whole reply), the prose inside reads exactly as it would without it, and nothing you haven't been told about ever gets a box, since the box itself would give it away.

Each marker closes with its slash twin (%%/trance%%). The model only sees these instructions for the modes you picked.

After each reply, a quick rule pass flags suspect lines, and the Editor model rewrites only those lines; nothing else in the reply changes. ⟲ in a message's menu switches between the original and the fixed version.

Add your own phrases to the slop list, one per line. Run Editor on last reply runs it on demand, even while it's off.

Both live in the wand menu.

• Make card turns a finished Metaphysics Card Builder chat into a real character, with its Codex modes already picked.
• Check card reviews the open character's card: persona names where {{user}} belongs, a system prompt that replaces your preset's, too many tokens sent every reply, mind control written as a trapped victim, pointers in the greetings, and more. Deeper review asks your Notes model for a closer read.

Background jobs run through your connection profiles, separate from the model you chat with. Use a local model or a cheap, fast API model for all three, not your roleplay model: the jobs are small, so a budget model does them well, and they stay quick and nearly free. A job with no model picked simply doesn't run.

All display-only: they change how messages look, never what the model reads.

You can edit, disable or delete them; the extension leaves your changes alone. They're also on the site as files, in case you ever need to import them yourself.

The extension sends chat text only to the connection profiles you pick under Models: the Ledger model reads the last few messages after each reply, the Editor model reads the lines it fixes, the Notes model reads the messages around an ended body change, and a deeper card review sends that card. Nothing else leaves your SillyTavern, and there's no telemetry.

• Nothing seems to happen. Pick the character's modes (✏️ under Modes), and make sure a model is picked for Ledger & suggestions.
• The Ledger got something wrong. Tap ✏️ on the Body Ledger and fix it; the next update builds on your fix. 🔄 rebuilds it from the card and the last few messages.
• Deeper review takes a while. A thorough read usually takes under a minute on a cheap model; the button counts the seconds, a second tap cancels, and after three minutes it stops on its own.
• I see %%trance%% or other markers as text. The display scripts are off or gone: check Extensions → Regex, or import them from the site.
• "This version of Worldhopper has a newer Codex". Your copy of WH Metaphysics has edits of your own, so it wasn't replaced. Update (keep mine as a copy) installs the new one and keeps yours alongside.
• The Ledger stays off. With only Timestop picked there's nothing for it to track; otherwise check that it's on and has a model.
• Text Completion. The Codex and the Ledger work the same; the Editor's style fixes skip Text Completion replies unless you turn off Chat Completion replies only.

extras/wh-power is a small Windows server plugin that adds Shut down SillyTavern to the wand menu, for switching SillyTavern off from your phone. It isn't installed with the extension; see its README.

Worldhopper Engine by Coz, built with Claude. The Worldhopper 1.0 presets, the Metaphysics Card Builder and the WH Cardmaker preset are on coz-rp.github.io.