System-Wide Local AI Text Injector

Native-feeling macOS menu bar utility (Python + rumps) that rewrites copied text using a local Ollama model, then pastes the result back into the previously focused app.

Current Status

• V1 is working for plain-text workflows.
• Target flow: copy text (Cmd + C) -> click menu action -> AI rewrite -> auto paste (Cmd + V).
• Designed for local model testing and prompt iteration.

Features

• Menu bar app with 4 actions:
  • Fix Grammar & Spelling
  • Make Professional
  • Build Prompt
  • Draft Reply
• Local Ollama inference via http://localhost:11434/api/generate
• Clipboard snapshot + auto-restore safety
• Re-focus previous app and auto-inject via paste
• Undo last injection
• Terminal debug logs for:
  • active model and source
  • copied text
  • generated prompt
  • raw model output
  • cleaned output used for paste
• Refusal-recovery retry path for smaller models

Requirements

• macOS
• Python 3.11+
• Ollama installed and running
• A local model pulled (example):

ollama pull llama3.2:1b

Quick Start

python3 -m venv .venv
source .venv/bin/activate
pip install -e .
python main.py

Alternative run command:

local-ai-text-injector

Permissions (macOS)

Grant permissions to the host process running the app (usually Terminal or iTerm):

• Accessibility
• Input Monitoring (if prompted)

Without these, focus switching and paste injection will fail.

Model Selection

You can switch models without code edits.

Option 1: edit model.txt:

llama3.2:1b

Option 2: use environment variable:

export OLLAMA_MODEL=qwen2.5:3b

Selection priority:

1. OLLAMA_MODEL
2. OLLAMA_MODEL_FILE
3. ./model.txt
4. project model.txt
5. fallback llama3.2:1b

Note:

• If OLLAMA_MODEL is set, model.txt changes are ignored.
• Model selection is re-read each time you click an action.

Environment Variables

• OLLAMA_BASE_URL default: http://localhost:11434
• OLLAMA_MODEL optional override (highest priority)
• OLLAMA_MODEL_FILE optional custom model file path
• OLLAMA_TIMEOUT_SECONDS default: 45
• CLIPBOARD_RESTORE_DELAY_SECONDS default: 1.25
• TERMINAL_DEBUG_LOG default: true

Typical Usage

1. Select text in any app.
2. Press Cmd + C.
3. Click the menu bar icon ✍️.
4. Choose an action.
5. Wait for success notification.
6. Result is pasted into the previous app.

Project Structure

• main.py entrypoint
• src/text_injector/menu_app.py menu bar UI and action wiring
• src/text_injector/workflow.py orchestration logic
• src/text_injector/actions.py action prompts and prompt builders
• src/text_injector/ollama_client.py Ollama API client
• src/text_injector/clipboard.py clipboard snapshot/read/write/restore
• src/text_injector/injector.py focus and paste automation
• src/text_injector/settings.py env/file config resolution
• model.txt quick model-switch file

Limitations (V1)

• Plain text only (rich text not preserved)
• Assumes user copied text before action
• UI automation depends on macOS permissions
• Small models may still produce weak rewrites on complex input

Troubleshooting

• No paste happened:
  • Check Accessibility/Input Monitoring permissions.
• Model didn’t switch:
  • Confirm terminal log lines Model: and Model Source:.
  • Run unset OLLAMA_MODEL if you want model.txt to control model.
• Output quality is poor:
  • Try a larger model (llama3.2:3b, qwen2.5:3b, etc.).
• App seems unchanged after edits:
  • Restart process with Ctrl + C then python main.py.

License

Add your preferred license before public release.