pixlstash-lmstudio

An LM Studio plugin that illustrates the model's chat replies with pictures from a PixlStash server. The model decides when to illustrate by calling a single tool; the plugin runs the semantic search, downloads a matching picture, and renders it inline in the assistant's reply.

How it works

• Registers as an LM Studio prediction-loop handler that drives the user-selected model via model.act() with one tool: show_pixlstash_image.
• The model composes a vivid scene/character query as the tool argument.
• The plugin searches PixlStash (GET /api/v1/pictures/search), filters to renderable formats, picks one of the top-N results that hasn't already been shown in this chat, downloads it into the chat working directory, optionally downscales it to a max height, and writes a markdown image into its own assistant content block.
• The picture renders inline because LM Studio displays markdown images whose src is a local file in the chat working directory.
• A per-chat "Max images per response" cap and a chat-wide no-duplicates rule prevent runaway tool calls.

The model never has to echo any markdown — the plugin renders the image itself — so this works with roleplay fine-tunes that won't repeat tool results.

More examples

Install

Prerequisites

• LM Studio (the lms CLI ships with the app — make sure it's on your PATH, usually at ~/.lmstudio/bin).
• A PixlStash server you have a READ-scoped share token for (PixlStash → User Settings → Tokens).

From the LM Studio Hub (end users)

Browse to pixlstash-lmstudio on the LM Studio Hub, or pull it via the CLI:

The plugin will appear in LM Studio's plugin list. Open the plugin's gear icon and fill in the PixlStash base URL + your READ share token (and turn on Ignore HTTPS certificate errors if your PixlStash uses a self-signed cert). That's it — start a chat and the model can now call show_pixlstash_image.

From source

Running the plugin from source, the project layout, and publishing instructions live in CONTRIBUTING.md.

Configure

Global (plugin gear icon, shared across all chats):

Setting	Default	What it does
PixlStash base URL	http://localhost:9537	Where PixlStash lives. Use https://… for TLS.
READ share token	(empty)	A READ-scoped token from PixlStash → User Settings → Tokens.
Ignore HTTPS certificate errors	off	Accept self-signed certs (local servers only).

Per chat (per-conversation settings panel):

Setting	Default	What it does
Pick from top N results	10	Random pool size for each search.
Minimum semantic score	0.5	PixlStash search threshold.
Image to inject	Full	Full image vs. 256×256 thumbnail.
Max image height (px)	512	Downscale taller full-mode images (0 = no resize).
Max images per response	1	Hard cap on tool calls per reply (0 = unlimited).
Allowed formats	jpg/jpeg/png/webp	Restrict to LM Studio-renderable formats.
Required / excluded tags	—	Tag-based include/exclude filters.
Minimum user score	0	PixlStash per-picture rating floor.
Limit to characters / sets / project	—	Scope the search to specific PixlStash entities by name.

System-prompt hint

The plugin works best when your system prompt mentions the tool — for example:

When a visual would help illustrate a scene, character, or object, call show_pixlstash_image with a short, vivid description. You don't need to repeat any markdown; just narrate around it.

Contributing

Bug reports, feature requests, and pull requests welcome — see CONTRIBUTING.md for the dev setup, project layout, and publishing workflow.

License

MIT © Pikselkroken