A production-grade personal assistant that runs entirely on your own hardware — not a toy. The E*TRADE integration and code development toolset fill real gaps left by commercial AI tools: full portfolio analysis with real financial data, AI-driven coding with git integration and diff previews, and complete privacy with no cloud dependencies. Built for and tested with Qwen3.5-35B-A3B on an NVIDIA RTX 5090. Useful today for automating portfolio analysis, developing software, or running local LLM experiments.
Connect any local LLM via llama.cpp and get a full-featured chat interface with web search, code execution, interactive visualizations, AI-powered code development, travel planning, and deep E*TRADE brokerage integration.
git clone https://github.com/ollls/ScrapChat.git
cd ScrapChat
npm install
cp .env.example .env # Edit with your settings
npm run css:build # Build Tailwind CSS
npm start # Open http://localhost:3000Requires: Node.js >= 20 and a running llama.cpp server (see setup below).
ScrapChat connects to a local llama.cpp server. Tested and running great on an NVIDIA RTX 5090 with the Qwen3.5-35B-A3B mixture-of-experts model (only 3B active parameters — fast inference with strong reasoning at 131K context).
#!/bin/bash
# run-qwen-server-5090.sh
export CUDA_VISIBLE_DEVICES=0
./llama.cpp/build/bin/llama-server \
-hf unsloth/Qwen3.5-35B-A3B-GGUF:UD-Q4_K_XL \
--ctx-size 131072 \
--gpu-layers 99 \
--flash-attn on \
-t 16 \
--temp 1.0 \
--top-p 0.95 \
--min-p 0.01 \
--top-k 40 \
--repeat-penalty 1.05 \
--cache-type-k q8_0 \
--cache-type-v q8_0 \
--mmap \
--host 0.0.0.0 \
--port 8080Why this works well:
--gpu-layers 99— full GPU offload, the 5090 has plenty of VRAM--flash-attn on— flash attention enables the full 131K context window--cache-type-k/v q8_0— quantized KV cache fits 131K context in VRAM- Qwen3.5-35B-A3B is a MoE model — only 3B params active per token, so inference is fast while total knowledge spans 35B
Any llama.cpp-compatible model works. Adjust --ctx-size and quantization for your GPU. Smaller models like Qwen3-8B or Llama-3.1-8B run fine on GPUs with less VRAM.
ScrapChat is a universal assistant. Ask it anything — it picks the right tools automatically:
- Search the web and summarize articles
- Check weather and plan trips with hotel search and booking
- Write and run Python scripts for data analysis, charts, and reports
- Execute shell commands on your machine
- Generate interactive dashboards — Chart.js, SVG, and HTML visualizations right in chat
- Manage your E*TRADE portfolio — holdings, options, transactions, real-time quotes
- Develop code on any project — read, edit, write, delete source files with diff previews, git integration, and test runner
- Pin conversations — keep important chats across server restarts
Point the AI at any project directory and it becomes a coding assistant. It can browse your codebase, make targeted edits with color-coded diff previews, manage git, and run your test suite — all from the chat interface.
- Read & search — browse file tree, read files, regex search across the codebase
- Edit code — exact string replacement with uniqueness checks and whitespace tolerance. Every change shows a diff preview (green = added, red = removed) before applying
- Write & delete — create new files or remove unused ones during refactors
- Git integration — status, diff, log, commit, branch, push — with safety tiers that block destructive operations (force push, reset --hard, rebase)
- Run tests — configurable per project (
npm test,pytest,cargo test, etc.) — the AI runs tests after changes and fixes failures automatically - Switch projects — tell the AI "work on ~/prj/other-project" and all source tools retarget (always requires approval)
Works on any language or framework. Set SOURCE_DIR in .env to get started.
The LLM doesn't do math — Python does. When you ask "what's my portfolio beta?", the LLM fetches your holdings via E*TRADE, writes a Python script with pandas/numpy, runs it, and renders the result as an interactive chart. Full precision, no hallucinated numbers.
- Portfolio allocation, P&L, unrealized gains
- Option chains with full Greeks (Delta, Gamma, Theta, Vega, IV)
- Transaction history with auto-pagination
- Covered call screening, tax-loss harvesting, performance tracking
Seven colored buttons in the top bar — each represents a session type. Click one to start a new chat. Save a session prompt per color and it auto-submits on creation. Hover any button to see its saved prompt title.
Examples: a daily briefing session, a coding assistant session, a financial analysis session, a support/help session.
A library of reusable text snippets. Click to load into the input box. Supports variables that prompt you for input:
{$date},{$time},{$location}— auto-filled{$City},{$Symbol}— text input dialog{$CheckIn:date},{$Stay:daterange}— calendar pickers
Save any input text as a prompt with one click — title is auto-generated.
Save any visualization the AI creates and reuse it. Type [template: Weather] and the AI regenerates the same dashboard layout with fresh data. Great for recurring reports and dashboards.
All three menus support drag-to-reorder and inline title editing.
All settings via .env:
| Variable | Description |
|---|---|
PORT |
Web server port (default: 3000) |
LLAMA_URL |
llama.cpp server URL (default: http://localhost:8080) |
LOCATION |
Your default location for weather/travel (e.g. "Oakland Park, FL") |
SOURCE_DIR |
Project root path — enables AI code development tools |
SOURCE_TEST |
Test command (e.g. npm test, pytest, cargo test) |
PYTHON_VENV |
Path to Python venv for code execution |
SEARCH_ENGINE |
keiro, tavily, or both |
TAVILY_API_KEY |
Tavily search key |
KEIRO_API_KEY |
Keiro search key |
LITEAPI_KEY |
LiteAPI key for hotels/travel (optional) |
ETRADE_CONSUMER_KEY |
E*TRADE OAuth key (optional) |
ETRADE_CONSUMER_SECRET |
E*TRADE OAuth secret |
ETRADE_SANDBOX |
true for sandbox mode |
- Register at developer.etrade.com
- Add credentials to
.env - Click the E*TRADE indicator in the app and complete OAuth
Tokens are in-memory only — nothing persisted to disk.
python -m venv ~/finance_venv
source ~/finance_venv/bin/activate
pip install pandas numpy matplotlib scipy plotlySet PYTHON_VENV=~/finance_venv in .env.
- Top bar — Session buttons, service status indicators, LLM/search engine switcher, context usage, elapsed timer
- Sidebar — Conversation list with switching, delete, and pin (📌) to persist across restarts
- Chat — Markdown, syntax highlighting, Mermaid diagrams, collapsible reasoning, interactive applets
- Input — Image attachments (paste, drag, button), checkboxes for Applets/Autorun/Think
- Diff previews — Color-coded diffs for all source code changes, visible even with Autorun enabled
- Menus — Tools (toggle on/off), Prompts, Sessions, Templates
The AI has 20 built-in tools it uses automatically:
| Tool | What it does |
|---|---|
web_search |
Search the web |
web_fetch |
Read a webpage |
run_python |
Execute Python scripts |
run_command |
Run shell commands |
save_file / list_files / file_read |
File management |
source_project |
Switch to a different project directory |
source_read |
Browse source code (tree, read, grep) |
source_write |
Create or overwrite source files |
source_edit |
Targeted edits with diff preview |
source_delete |
Remove source files |
source_git |
Git with safety tiers (blocks destructive ops) |
source_test |
Run project tests (configurable per project) |
etrade_account |
E*TRADE portfolio, quotes, options, orders |
hotel / travel / booking |
Hotel search, weather, trip booking |
current_datetime |
Current time and timezone |
Tools can be toggled on/off at runtime from the Tools menu. Add your own in src/services/tools.js.
npm run dev # Auto-reload server
npm run css:watch # Watch Tailwind CSS (separate terminal)Node.js, Express v5, Tailwind CSS v4, vanilla JS frontend. No bundler, no framework, no build step beyond CSS.
ISC