InvisibleGPT: Ambient AI Assistant

A minimalist tool for capturing prompts and delivering AI responses through various input/output methods, designed for zero context switching.

What I Built

Three Input Modes:

• Clipboard: Automatically processes when you copy text
• Keyboard: Cmd+. to capture, Cmd+Shift+. to submit
• Terminal: Direct CLI input for scripting

Four Output Modes:

• PDF: Two-column research paper format via pdflatex
• Clipboard: Silent copy, paste anywhere
• Terminal: Colored output for readability
• File: Append to log file

Plus:

• Conversation memory across interactions
• Configurable model selection (o4-mini, gpt-4o, gpt-3.5-turbo, etc.)
• TOML-based configuration with environment variable overrides

Technical Details

Configuration System

[model]
name = "o4-mini-2025-04-16"

[input]
method = "keyboard"

[output]
method = "clipboard"
file = "output.txt"

[pdf]
pdflatex_path = ""
output_file = "output.pdf"

Override via CLI:

python src/main.py --input clipboard --output pdf --model gpt-4o

LaTeX Character Escaping

Automatic handling of special characters (\, {, }, $, &, %, #, _, ~, ^) for PDF generation. Prevents most formatting errors.

Clipboard Polling

last_clipboard = ""
while True:
    current = pyperclip.paste()
    if current != last_clipboard and should_process(current):
        process_prompt(current)
    time.sleep(0.1)  # 100ms poll

PDF Pipeline

LaTeX source → pdflatex → cleanup aux files → output PDF. Two-column article format with proper typography.

Use Cases

Research Writing: Copy question → PDF appears. No app switching.

Coding: Hit shortcut → ask question → response in clipboard → paste into code.

Batch Processing: Script feeds prompts via terminal → all responses append to single file.

Performance

• Clipboard detection: <100ms
• API request: 1-5s depending on model
• PDF generation: 500ms-1s
• Memory: ~50MB
• CPU (idle): <1%

Learnings

What Worked:

• Zero context switching with clipboard/keyboard modes
• TOML config is cleaner than JSON
• LaTeX escaping prevented 90% of PDF failures
• Different output modes for different workflows

What Could Be Better:

• Clipboard polling wastes CPU; OS-level hooks would be better
• No streaming (waits for full response)
• Single conversation only (no session management)
• LaTeX template is inflexible

Unexpected Use Cases:

• Live lecture note-taking (clipboard mode)
• Code review comments (keyboard → clipboard)
• SQL generation from natural language (terminal scripting)

Future Work

• Streaming responses for better UX
• Multiple conversation sessions
• Custom templates (markdown → PDF, HTML output)
• OS integration (macOS Services, Windows context menu)
• Smart context detection (code, equations, URLs)