A lightweight chat UI for running local language models via Ollama. Single HTML file, no frameworks, no cloud dependencies.

• Streaming chat with markdown, syntax highlighting, and LaTeX math
• Vision — upload images for models that support it (e.g. Qwen3.5, Gemma3)
• Agent tools — the model can autonomously use:
  • Web search — DuckDuckGo, model decides when to search
  • Browse URL — fetch and read any web page
  • Calculator — evaluate math expressions (supports Python math functions)
  • Python runner — execute code with matplotlib plot support (inline charts)
  • Notes — save, list, and read local markdown notes
  • PDF generation — create printable documents from markdown
• PDF upload — extract and summarise text from uploaded PDFs
• Thinking — toggle visible chain-of-thought reasoning
• Tool memory — tool calls and results persist across turns (collapsed, expandable in chat)
• Context counter — real token count from Ollama (warns at 70%, critical at 90%)
• Print chat — full conversation with expanded thinking blocks
• Configurable — model, temperature, context length, max tokens, system prompt

• Ollama installed and running
• Python 3.x
• pdfplumber and matplotlib for PDF extraction and plotting:

  pip3 install pdfplumber matplotlib

• At least one model pulled, e.g.:

  ollama pull qwen3.5:9b-q8_0

# 1. Start Ollama (if not already running)
OLLAMA_KEEP_ALIVE=-1 ollama serve

# 2. Start the chat server
cd local_llm
python3 serve_ui.py &

# 3. Open in browser
open http://127.0.0.1:3000/

local_llm/
  chat-ui.html    # Complete chat UI (single file)
  serve_ui.py     # Proxy server + agent tool endpoints
  vendor/         # Bundled JS/CSS (marked, highlight.js, KaTeX)
  notes/          # Saved notes (created at runtime)
  generated/      # Generated PDF pages (created at runtime)

Browser (:3000)  <-->  serve_ui.py  <-->  Ollama (:11434)
                          |
                      /search    (DuckDuckGo)
                      /browse    (fetch web pages)
                      /calc      (math expressions)
                      /run_code  (Python + matplotlib)
                      /save_note, /list_notes, /read_note
                      /extract_pdf, /generate_pdf

• serve_ui.py serves the UI and proxies /api/* requests to Ollama
• Agent tool endpoints handle search, browsing, code execution, notes, and PDFs
• Tool calling is automatic — the model decides when to use tools
• All processing runs locally — no data leaves your machine (except search queries and browsed URLs)

• 32GB RAM + large model: set Context Length to 2048 in settings to avoid memory pressure
• Thinking toggle: turn off for faster responses when you don't need reasoning
• Vision: attach images via the paperclip button or drag & drop
• Search: the model automatically uses web search when it needs current information
• Plots: ask the model to create charts — matplotlib runs locally, plots display inline
• Context counter: shows real token usage from Ollama — start a new chat when it gets high
• Models without tool support (Gemma3, DeepSeek): tools are automatically disabled