Endpoint Settings

🔌 Endpoint Settings

The Endpoints tab in Settings sets which AI the app connects to. It has its own tabs: Text, Image and, in Advanced mode, Tag Prompt.

• To set up a text endpoint step by step, follow Connect Your Own AI.
• To set up an image server, follow Image Generation.

In the tables below, Advanced marks a row that shows in Advanced mode only. See Simple and Advanced.

Text

The Preset list holds your saved endpoints:

• Demo AI is the shared cloud endpoint. You can't edit it.
• Formamorph AI is the preset for Patreon Supporters. It shows your allowance in place of the fields. See Formamorph AI for Supporters.
• On the desktop app, Built-In Engine runs a model on your PC.

Add New Preset… makes a new preset. Duplicate copies Demo AI or a preset you made. Rename, Reset and Delete act on a preset you made.

Setting	What it does
Endpoint URL	The address of your model server. When the app completes the address, the line under it shows the full URL that requests go to. Trouble Connecting? opens a checklist.
API Token	The token for a hosted service. Leave it empty for a local server.
Model Name	The model the endpoint uses, exactly as the server names it
Context Window	Advanced. How many tokens the model keeps in context. Detect asks the server.
Max Output Tokens	Advanced. Caps how long each answer can be. It does not cap reasoning. To set it, select Override Endpoint Limit. Without it, the cap is No Limit.
Sampling	Advanced. Five rows: Temperature, Repetition Penalty, Top-p, Top-k and Min-p. Each row has a checkbox, and an unchecked row sends nothing. For Temperature and Repetition Penalty, a prompt's own value and a built-in prompt's value win over this row.
Extra Parameters	Advanced. A JSON object added to every request to this endpoint, such as a vendor's reasoning switch, top_a, chat_template_kwargs or a routing block. See Extra Parameters.

Reset AI Endpoint, at the bottom of the tab, sets the URL, model name, token and limits back to their defaults. It also clears Extra Parameters.

Extra Parameters

• These fields only add what the app leaves out. A field the app sets for a request wins over the same field here.
• You can't set model, messages, stream, tools, tool_choice, response_format, n or stop.
• Text with an error shows why in red. Requests then go out without these fields.
• Where the app sends a reasoning field, the Native Reasoning row and each prompt's own setting control it. See Reasoning.

Built-In Engine

On the desktop app, the Built-In Engine preset shows the engine's own panel in place of the fields. The panel has its own Simple and Advanced switch. Every row below Local Model is in the Engine section. See The Desktop Engine.

Setting	What it does
Local Model	Manage Models… opens the Local model dialog to download and load a model. Its Options view holds Auto-Load, on by default, which loads a model as soon as the engine is in use.
Context Size	How much recent story the model can see. More context uses more GPU memory. Lower it first when a model does not fit.
GPU	Simple on the panel. Runs the model on the GPU. Off runs it on the CPU, which is slower.
GPU Layers, Layers	Advanced on the panel. Auto puts as many layers on the GPU as fit. Max puts all of them there and can run out of memory. Custom sets the count in Layers.
GPU Device	Which GPU loads the model. Shows while the GPU is in use. See the list below.
Temperature	How much randomness the model uses when it picks each word. About 0.7 fits most story models.
Max Output Tokens	Caps how long each answer can be. An answer at the cap ends at its last full sentence.
Top-p, Top-k, Min-p, Repetition Penalty	Advanced on the panel. Sampling limits on which words the model picks.

GPU Device has these choices:

• Auto uses your discrete GPU. With more than one, it uses all of them. With none, it uses the integrated GPU.
• All GPUs always splits a model across every GPU.
• A GPU by name uses only that GPU.

When you change the panel:

• Sampling changes apply on the next turn.
• The other changes need Save & Reload Model.
• Reset to Defaults sets the panel back.

Engine Memory and Context

These rows are on the Built-In Engine panel. They show when the panel's own switch is on Advanced.

Setting	What it does
Parallel Requests	How many requests the model answers at the same time. With Unified KV Cache on, they share one context. With it off, each gets the full Context Size and uses more GPU memory.
Unified KV Cache	On by default. Parallel requests share one context, so extra requests add no GPU memory. A long request leaves less room for the others. Turn it off to give each request the full Context Size.
K Cache Type, V Cache Type	F16 is full precision. Q8_0 and Q4_0 shrink the context, so a longer Context Size fits, at a small cost in quality. A quantized V Cache Type needs Flash Attention. While Flash Attention is off, V Cache Type can't be Q8_0 or Q4_0.
KV Cache in RAM	Keeps the context in system RAM to free GPU memory. Replies get much slower, and the slowdown grows as the context fills. Try a smaller K Cache Type and V Cache Type first. Use this only when the model still does not fit.

Flash Attention

Flash Attention is Advanced on the panel, and on by default. It uses less GPU memory and is often faster. Turn it off when an old GPU can't run it. A quantized V Cache Type needs it on.

Context Overflow

Context Overflow is Advanced on the panel. It sets what happens when the story is longer than Context Size:

• Truncate Middle (default) keeps the system prompt and your first message, and drops messages from the middle.
• Rolling Window keeps the system prompt and drops the oldest messages first.
• Stop at Limit stops the request with an error.

It never cuts a message in half. A change applies on the next turn, with no reload.

Image

Setting	What it does
Preset	Your saved image endpoints. You can edit each one, Default too.
Enable Image Generation	On by default. Shows the Generate with AI buttons. Off hides the other rows of this tab, the Tag Prompt tab and Scene Images.

The Connection section sets which server makes the images. The setup steps for each provider are in Image Generation.

Setting	What it does
Provider	ComfyUI, InvokeAI, Automatic1111 / Forge, NovelAI, or an OpenAI-compatible service on the desktop app. How to Set Up shows the steps for the picked provider.
Endpoint URL, API Token	The image server address and its token
Model	The checkpoint that makes the image. On ComfyUI it fills the workflow's %ckpt%. A workflow without %ckpt% loads its own model, and the row shows that model, grayed out.

The Image section sets how each image is made.

Setting	What it does
Prompt Prefix	Quality and style tags put before every image prompt
Negative Prompt	Tags the image must not have
Clean Up Tags	On by default. Rewrites scene and generated tags as plain lowercase words, without weights or other prompt syntax. Turn it off to send your entity and location tags as you wrote them. It never changes Prompt Prefix or Negative Prompt.
Portrait Size, Landscape Size	Advanced. Image width and height. Portraits are for entities. Landscapes are for locations and thumbnails.
Steps / CFG	Sampling steps, and how closely the image follows the prompt
Sampler	The sampling method
Face Fix	Automatic1111 / Forge and InvokeAI only. Draws faces again in a second pass. On Automatic1111 / Forge it needs the ADetailer extension on your server. On InvokeAI it about doubles the time.
Workflow	Advanced, ComfyUI only. Replaces the default ComfyUI graph with one exported in API format. How to Get This shows how to export one. Reset to Defaults brings back the default graph.
File Name	Advanced, ComfyUI only. Where each image is saved under ComfyUI's output folder. See File Name.
Board	Advanced, InvokeAI only. The InvokeAI board that gets the images
Qwen3 Encoder, Z-Image VAE, Anima VAE	Advanced, InvokeAI only. Shows for a Z-Image or Anima model. Leave them empty to pick one automatically.

File Name

• The pattern takes {world}, {subject}, {turn} and {preset}. {subject} is an entity's or location's name, or turn-12-scene.
• The default is formamorph/{world}/{subject}.
• A / in the pattern makes a folder.
• A custom workflow uses it where its save node's prefix is %prefix%. Without one, the row is grayed out.

Tag Prompt

Advanced. Shows when Enable Image Generation is on. It is the prompt that your text model gets to turn a description into image tags. Reset to Defaults brings back the original text.