GLM Studio

Run GLM (and any Ollama model) locally, in a web UI that's actually nice to use.

🖼️ Hero screenshot coming. Capture the app with a slide deck rendered in the artifact panel and the Export → .pptx button visible — easiest is to run python3 server.py, open http://localhost:7860/?demo (it seeds a sample deck), and screenshot. Save it as docs/screenshot.png (or docs/demo.gif for a short animation), then replace this whole block with:

![GLM Studio — a slide deck rendered in the artifact panel](docs/screenshot.png)

GLM Studio is a self-hosted web interface for chatting with a local GLM model through Ollama — no cloud, no API keys, no monthly bill, nothing leaving your machine. It works with anything Ollama can run (glm-5.2:cloud, llama3, qwen, etc.), but it's built around GLM and tuned to make the most of what that family is good at: long-form, structured, content-shaped output.

Try it (two ways, no install of this app needed)

1. In your browser, against your own Ollama. Open the live demo, click the gear → Ollama URL, and paste http://localhost:11434. The page calls your Ollama directly — nothing in between. Ollama just needs to allow web pages to reach it, so start it once with CORS on:

OLLAMA_ORIGINS=* ollama serve

(The demo loads with a sample slide deck so you can see the artifact panel render even before you connect. Works in Chrome/Edge/Firefox; http://localhost is allowed from the HTTPS page. For an Ollama on another machine over plain http://, use the launcher below — the browser blocks cross-machine http:// from an HTTPS page.)

2. One click, fully local. If you've cloned the repo, double-click the launcher for your OS — it sets OLLAMA_ORIGINS=*, starts Ollama and the UI, and opens the browser. No terminal, no CORS to think about:

• macOS: start.command
• Windows: start.bat
• Linux: ./start.sh

Why this UI

Ollama ships with a basic terminal prompt. That's fine for a quick question, but the moment you ask a model for something richer — a deck, a chart, a set of flashcards, a report — raw text in a terminal is the wrong tool. GLM Studio gives that output a real home:

• An artifact panel that detects what the model produced and renders it properly — charts plot, spreadsheets become editable grids, slide decks paginate with speaker notes, flashcards flip, tables and documents get formatted. You don't read a wall of markdown, you get the thing.
• One-click export to real files — slide decks to a native, editable .pptx (with live PowerPoint charts), spreadsheets to .xlsx, tables to .csv, documents to PDF. The model's output becomes a file you can actually hand to someone.
• A content creator — pick a target (notes, study guide, flashcards, slide deck, comparison table, timeline, chart, summary) and the UI steers the model toward that format, so you get a usable artifact instead of a generic essay.
• A clean, fast, single-file app — no build step, no frameworks, no node_modules. One index.html plus a ~140-line Python proxy. Clone it, run it, you're in.

Think of it as a lightweight, local-first alternative to a hosted AI writing deck — but it runs on your laptop, talks to your own model, and exports files you own.

How it compares

• vs the ollama terminal prompt — Ollama's CLI is fine for a one-line question, but the moment you ask for a deck, a chart or a report, you're staring at raw text. GLM Studio renders that output (charts plot, decks paginate, flashcards flip) and exports it to real files.
• vs Open WebUI — Open WebUI is a full server app with a database, users and a container stack. GLM Studio is one HTML file and ~140 lines of Python. No accounts, no Postgres, no Docker required to try it. The trade-off: it's a single-user, single-machine tool by design, not a multi-user server.
• vs a hosted AI deck/writing tool — no cloud, no API key, no bill, nothing leaves your machine. You also own the model and the output files.

Quick start

Easiest: double-click start.command / start.bat / start.sh (see above) — it does this for you. Manually:

python3 server.py
# open http://localhost:7860

That's it. The server serves the UI and proxies /api/* calls to Ollama at http://localhost:11434 by default.

You'll need Ollama running with a model pulled, e.g.:

ollama run glm-5.2:cloud

Point at a different Ollama host with the OLLAMA_HOST env var:

OLLAMA_HOST=http://192.168.1.50:11434 python3 server.py

Run with Docker

docker compose up -d
# open http://localhost:7860

The container starts the same proxy and expects Ollama reachable at http://host.docker.internal:11434 (override with OLLAMA_HOST in compose.yaml).

Settings

Click the gear icon in the top bar to set:

• Model — which Ollama model to talk to (default glm-5.2:cloud)
• Custom system prompt — override the built-in default for steering tone/role

Features

Chat

• Streaming replies, copy / edit / regenerate / retry
• Conversation history, pinning and archive — all stored in localStorage, nothing leaves the browser except the calls to your Ollama instance
• Drag-and-drop file and image uploads; images are auto-downscaled before being sent so requests stay inside the model's context window
• KaTeX-rendered math, syntax-highlighted code blocks

Artifact panel

When the model output contains a fenced block or a slide deck, GLM Studio detects it and renders it in a side panel instead of leaving it as raw text:

Format	Detected from	Export
Chart	```chart JSON spec	— (rendered inline)
Spreadsheet	```spreadsheet	.xlsx
Table	markdown table	.csv, markdown
Slide deck	## Slide N: <layout>	.pptx (native charts), HTML, PDF
Flashcards	```flashcards JSON	.json, .csv
Document / report	prose with headings	markdown, HTML, PDF
Code	fenced code block	copy, raw

Content creator

The ✨ button in the composer opens a mode picker that steers the model toward a target artifact — useful when you want a specific output rather than a free-form answer:

• Notes · Study guide · Flashcards · Slide deck · Comparison table · Timeline · Chart · Summary

Slide decks

Slide decks use a small text format (## Slide N: <layout> with title, bullets, split, chart, quote layouts) plus optional speaker notes via a trailing Notes: line. Export to a real, editable .pptx with native PowerPoint charts (bars, lines, pies, doughnuts) via PptxGenJS — fully client-side, the deck never leaves the browser.

Examples

A real prompt and what comes back — the repo ships a sample deck (indian-tv-slides.md) you can open in the app, or see rendered in the live demo.

Prompt: Make me a slide deck on the evolution of Indian television, from Doordarshan to OTT.

The model returns a slide_deck artifact the panel renders straight away — title, bullets, a chart slide, a split comparison, a quote with speaker notes — and the Export → .pptx button hands you a native PowerPoint file with the chart live and editable.

A second prompt — "now make flashcards from the key facts" — returns a flashcards artifact that renders as a flippable card deck you can click through and shuffle, and export to .csv / .json.

Other things that work well: "compare these three phones as a table", "make a study guide on the French Revolution", "plot quarterly revenue as a bar chart".

Roadmap

A few things worth doing next — feedback and PRs welcome:

• More slide layouts (image, two-column stats, agenda with timings)
• Markdown document export to .docx
• Conversation folders / tags
• Streaming artifact rendering (render as the model writes, not only at the end)

Project layout

index.html        # the whole UI — HTML, CSS and JS in one file
server.py         # static file server + /api/* proxy to Ollama
start.command / start.bat / start.sh   # one-click launchers (macOS / Windows / Linux)
Dockerfile
compose.yaml
*.md              # design docs that drove the build (ARCHITECTURE, DESIGN-SPEC, …)
indian-tv-slides.md, slides.json, view-slides.html   # a sample deck

Troubleshooting

• "Couldn't reach Ollama" on the hosted demo — Ollama isn't running with browser access. Start it with OLLAMA_ORIGINS=* ollama serve, then re-test the URL in Settings. Works for http://localhost in Chrome/Edge/Firefox; for an Ollama on another machine, use the local launcher (the browser blocks cross-machine http:// from an HTTPS page).
• HTTP 400 — prompt is too long — the conversation + images exceeded the model's context window. Start a new chat, or the downscaler should keep image attachments small.
• HTTP 502 / transient 5xx — the proxy retries once or twice; cloud models occasionally hiccup and a retry usually clears it.
• model not found, check the model name in Settings — the name in Settings doesn't match a model Ollama has pulled. Run ollama list and copy the exact tag.

License

MIT — see LICENSE.