Generate scripts
locally. Instantly.

A CLI agent that writes, validates, and self-corrects scripts using local LLMs. There when you need it. Works on ~4-8 GB RAM.

Learn More View on GitHub

bash - 80×24

$ scripy -p "rename all my jpegs by date taken"

▸ scripy v0.1.0 - qwen2.5-coder:7b

▸ generating ⚬⚬⚬⚬

✓ syntax valid

? run script to validate? [y/n/e/v/a] › y

▸ running sandbox...

✓ validation passed

? write to disk as rename_jpegs.py? [y/n/v] › y

✓ wrote rename_jpegs.py 2.1KB 4.3s

$ pip install scripy-cli

// how it works

The agentic loop

Describe the script you need. scripy handles the rest. Generation, validation, and self-correction all before asking to write anything to disk.

▸

Generate

You describe what the script should do. The local model writes it based on your prompt, language, and any existing file you pass with --input.

✓

Validate

scripy runs a syntax check (py_compile, bash -n) and optionally executes the script in a sandboxed subprocess with a configurable timeout.

Self-correct

If validation fails, the error is fed back to the model for up to 3 iterations. Each revision is diffed and previewed live in TUI mode.

Confirm & write

Before any side-effectful action - running or writing - scripy gates and asks for confirmation. Use -y to skip for non-interactive use.

// confirmation gates

You stay in control

scripy never runs or writes anything without asking. Every side-effectful action is gated by user input.

Run gate - before sandboxed execution

? run script to validate? [y/n/e/v/a] ›

y run it n skip e open in $EDITOR v preview script a always yes

Write gate - before writing to disk

? write to disk as dedup.py? [y/n/v] ›

y write file n skip write, print to stdout v preview file

Use -y / --yes to bypass all gates non-interactively.

// features

Everything you need,
nothing you don't

⚬ local-first

No cloud needed

Runs entirely on your machine via Ollama or LM Studio. Optional cloud integration via OpenAI is available too.

▸ agentic loop

Generate → validate → correct

Multi-turn self-correction loop feeds syntax errors and runtime failures back to the model to improve the script iteratively.

✓ sandboxed validation

Safe by default

Syntax is checked before execution. Sandbox runs use subprocess with a configurable timeout before you test it on your system.

~ TUI mode

Full interactive TUI

Launch with --tui for a split-pane interface: agent log on the left, live syntax-highlighted preview on the right with diff on revision.

? multi-language

Python, Bash, JS & more

Generate scripts in any language. Use --lang to specify or pick interactively in TUI mode.

▸ modify existing scripts

Refine scripts you already have

Pass an existing script with --input to modify, extend, or fix it while keeping the model grounded in the original.

// model support

Runs on small hardware

Designed for ~4–8 GB RAM. Tested with Ollama and LM Studio. Scales up to larger models with no config changes.

Model	Size	Tool calling	Code quality	Notes
llama3.1:8b	~4.7 GB	native	good	recommended
qwen2.5-coder:7b	~4.4 GB	inline	excellent	Best raw code quality
deepseek-coder:6.7b	~4.0 GB	inline	excellent	Strong Python support
llama3.2:3b	~2.0 GB	native	fair	Ultra-low resource fallback

Any OpenAI-compatible endpoint works. Set base_url in ~/.config/scripy/config.toml.

// examples

One command away

Just prompt from the command line. scripy figures out the rest.

bash

# Python

$ scripy -p "find duplicate files in a directory"

$ scripy -p "watch a directory for changes and notify me"

$ scripy -p "bulk resize images to 1080px" -o resize.py

# Bash / modify

$ scripy -p "backup home to /tmp" --lang bash

$ scripy -p "add a --dry-run flag" --input dedup.py

$ scripy -p "..." -y # skip all gates

Generate scripts locally. Instantly.