shellbud

sb suggests a shell command from a prompt.

Install

git clone https://source.tube/ilya/shellbud.git
cd shellbud
cargo install --path .

Usage

# TUI
sb

# CLI
sb "find files larger than 100 MB"

# Pipes work both ways
sb "what's eating my RAM?" | pbcopy
echo "weather in brooklyn" | sb

# Live dangerously
sb "find PDFs in my downloads folder" | bash

Configuration

# ~/.config/shellbud/config.toml
base_url = "https://openrouter.ai/api/v1" # OpenAI-compatible endpoint
api_key = "sk-or-..."
# Or get the api_key via a key manager command:
# api_key_cmd = ["op", "read", "op://Personal/OpenRouter/credential"]
model = "google/gemini-2.5-flash-lite"

Environment values override the config file:

• SHELLBUD_API_KEY or OPENROUTER_API_KEY — API key
• SHELLBUD_MODEL — model name
• SHELLBUD_BASE_URL — OpenAI-compatible API base URL

Development

make lint
make test
make eval

Benchmark results

To balance latency and accuracy, I chose Gemini 2.5 Flash Lite.

Model	Pass rate	Median E2E
GPT-5.4 Nano	30/50 (60%)	327 ms
Amazon Nova Micro	42/48 (87.5%)	516 ms
Gemini 2.5 Flash Lite	45/48 (93.8%)	521 ms
Ling 2.6 Flash	34/50 (68%)	532 ms
Gemini 3.1 Flash Lite	40/50 (80%)	546 ms
Amazon Nova Micro (Bedrock)	48/50 (96%)	746 ms
Mistral Small 2603	43/48 (89.6%)	769 ms
Qwen3 Coder Next	47/48 (97.9%)	804 ms
OpenAI GPT-4.1 mini	48/48 (100%)	925 ms
Claude Haiku 4.5	45/48 (93.8%)	1,035 ms
OpenAI GPT-4.1 mini	50/50 (100%)	1,313 ms
Claude Haiku 4.5	42/50 (84%)	1,357 ms
OpenAI GPT-4o mini	49/50 (98%)	1,382 ms