AI Porn Toolbox
18+
English
Submit a tool
Roleplay frontend · Updated October 15, 2026
SillyTavern logo

SillyTavern review: choose the backend before the setup

Choose SillyTavern if you want control over character prompts and model connections, and can maintain the setup. If you want a ready-to-chat service with a model included, its extra responsibility may outweigh its flexibility.

The interface is free and open source. You still need an inference backend. Running SillyTavern on your computer does not make a cloud-connected conversation local or remove the provider’s bill.

Official sources checked · Oct 15Latest stable checked · 1.19.0Research review · no score
Evidence: documentation and release notes

This revision checks official sources on October 15, 2026. It does not claim an installation, paid API session, network trace, latency measurement or model-quality test. The checks below are recommendations to run in your own setup.

Pick the model host before tuning the character

SillyTavern assembles the character, conversation and other prompt material for a connected model. The backend determines the model you can use; the frontend gives you a place to manage the conversation.

  1. 01

    Your browser

    The chat screen and controls you interact with.

  2. 02

    SillyTavern server

    Your running installation holds chat data and saved API keys.

  3. 03

    Inference endpoint

    A local server, cloud provider or Horde worker processes the prompt.

Follow the three layers. This is a simplified architecture guide, not a measurement of a running installation. Extensions can add other services.
Local inference

You run the model

Use a compatible model server on hardware you control. Model size and context drive its memory requirements; the frontend’s requirements do not size the model.

Choose it when: you want to manage both the model and its host.

Cloud API

A provider runs the model

You avoid hosting inference, but send prompt material to the chosen service. Check that provider’s price, retention and content rules before connecting.

Choose it when: you accept an external processor and its terms.

AI Horde

Volunteer workers run it

The documented out-of-box option uses volunteers’ GPUs. Available models, settings and wait times depend on workers; this is remote processing.

Choose it when: availability matters less than trying the connection.

Basis: what SillyTavern supplies, supported API connections and server-side data storage. The choice guidance is editorial analysis.

“Text Completion” does not mean “local”

Chat Completion and Text Completion describe how the request is formatted, not where it runs. Check the server URL and selected source. Choosing a completion type alone cannot establish privacy.

For a local host, see our existing KoboldCpp and Ollama profiles. These fill the model-serving role; SillyTavern can sit in front of a supported connection. Neither is required for a cloud-only setup.

Formatting distinction: official Chat Completion vs Text Completion explanation. Use the selected backend’s model-specific formatting instructions.

Save the preset, then save the connection profile

A connection profile remembers selections, not the unsaved contents of a preset. If you edit a template and switch profiles, those temporary edits can disappear. The profile name alone is a weak record of what will be sent.

A safer local-to-cloud switch

  1. In API Connections, choose the intended API, model and server URL.
  2. Save changed settings presets and formatting templates in their own controls.
  3. Use Create for a new connection profile, or Update to save changed selections to the current one.
  4. Use Information to inspect the saved profile. Before sending, verify the active source, model and server URL; start with a harmless prompt.

I would name profiles by destination and model, such as “Local — model name” and “Cloud — provider/model.” That makes a privacy or cost boundary easier to notice than names such as “Creative” and “Fast.”

Controls and save behavior: official Connection Profiles guide. The checklist and naming convention are suggested practice, not an observed test result.

Why an update matters here

The 1.19.0 release notes report a fix for profile switching that left the previous Chat Completion source active. Version 1.19.0 was the latest stable release checked on October 15. This describes a fixed issue, not a claim that current switching still misroutes requests.

A login does not encrypt stored chats or API keys

The multi-user documentation says chat history, API keys and other user data are stored in plain text on the SillyTavern server. Anyone with filesystem access can read or alter them. Account passwords separate users at the interface; they do not provide at-rest encryption.

On your own PC, that makes device access and backup handling part of the decision. On someone else’s host, it means trusting the operator with the stored files. A local-looking browser window does not settle either question.

Treat a downloaded data backup as sensitive as the original installation. Do not upload it to a public support thread to diagnose a character problem.

Storage and password limits: official multi-user documentation. Local inference also requires reviewing any extensions or other services you enable; it is not a privacy certification.

Back up the installation’s data, not just a character card

A character export is useful for moving a card. It does not replace the installation data needed to recover conversations, settings and connections.

For a modern Git or ZIP installation, keep the following map beside your backup. The official copy instructions below apply to versions 1.12.0 and later; older migrations and Docker have separate instructions.

data/
The user-data directory to preserve. For the default account and data path, user files live under data/default-user/.
config.yaml
The configuration file at the installation root. Keep it with the data when moving to a fresh install.
public/scripts/extensions/third-party/Only if applicable
Preserve this separately if you installed server-wide extensions for “All users.” It is not a reason to copy the entire old public/ folder.

For Windows, the update guide first points to UpdateAndStart.bat; Git installs can pull updates. For a ZIP replacement, unpack into a separate folder and transfer the documented data and configuration. Do not overwrite the new application with the whole old public/ folder.

The account menu also offers User settings → Account → Download Backup, which downloads that user’s data folder. That is useful, but it is not the same scope as preserving the server configuration.

File scope and update paths: official update guide. User backup control and account layout: multi-user guide. If you changed the data root or use containers, follow the matching documentation.

A phone browser still needs the server to be running

In the documented phone-to-PC arrangement, your PC runs SillyTavern and your phone opens its interface. Closing the server or putting the host to sleep interrupts that route; a browser shortcut does not turn it into a hosted service.

The server accepts only local-machine connections by default. Access from another device needs configuration and access controls. For access outside your network, the official guide recommends a VPN or tunnel rather than direct port forwarding.

Follow the official remote-connections guide for your arrangement. This review has not tested phone access, authentication or a remote deployment.

Next step: prove one model connection before building a library

I would choose the inference host first, then use a harmless prompt from SillyTavern’s Welcome Screen to check the connection. Its documentation describes this as a way to try a model without choosing a character first.

Only after that works, add one character and save one clearly named profile. Confirm you can reopen the conversation and locate a usable backup. If maintaining those pieces already feels burdensome, choose a hosted chat service instead.

Open the official setup documentation ↗