shell-ai 0.6.0

AI-assisted shell commands: create shell commands from natural language, or have them explained
shell-ai-0.6.0 is not a library.

Shell-AI

GitHub release Crates.io GitHub downloads Crates.io downloads Build status License: MIT Ko-fi

Describe what you want. Get shell commands. Or explain commands you don't understand.

What It Does

Suggest (shell-ai suggest or shai) turns natural language into executable shell commands. Describe what you want in any language, and Shell-AI generates options you can run, copy, or refine.

Explain (shell-ai explain) breaks down shell commands into understandable parts, citing relevant man pages where possible. Useful for understanding unfamiliar commands or documenting scripts.

Quick Start

# Install
cargo install shell-ai
ln -v -s shell-ai ~/.cargo/bin/shai  # Optional: shorthand alias for `shell-ai suggest`

# Configure
export SHAI_API_PROVIDER=openai
export OPENAI_API_KEY=sk-...

# Generate commands from natural language
shai "ファイルを日付順に並べる"  # Japanese: sort files by date

# Explain an existing command
shell-ai explain "tar -czvf archive.tar.gz /path/to/dir"

For guided configuration, run shell-ai config init to generate a documented config file.

Installation

[!TIP] After installing, configure your AI provider. Then, consider adding shell integrations for optional workflow enhancements.

From GitHub Releases

Download prebuilt binaries from the Releases page.

From crates.io

cargo install shell-ai
ln -v -s shell-ai ~/.cargo/bin/shai  # Optional: shorthand alias for `shell-ai suggest`

From Source

git clone https://github.com/Deltik/shell-ai
cd shell-ai
cargo install --path .
# Installs to ~/.cargo/bin/shell-ai
ln -v -s shell-ai ~/.cargo/bin/shai  # Optional: shorthand alias for `shell-ai suggest`

Features

  • Single binary: No Python, no runtime dependencies. Just one executable.
  • Shell integration: Tab completions, aliases, and Ctrl+G keybinding via shell-ai integration generate.
  • Multilingual: Describe tasks in any language the AI model understands. Responses adapt to your system locale.
  • Explain from man: shell-ai explain includes grounding from man pages, not just AI knowledge.
  • Multiple providers: OpenAI, Azure OpenAI, Anthropic, Claude Code CLI – plus OpenAI-compatible services (Groq, Ollama, Mistral) and Anthropic-compatible services (Ollama v0.14.0+).
  • Interactive workflow: Select a suggestion, then explain it, execute it, copy it, or revise it.
  • Live streaming previews: Watch suggestions and explanations generate token-by-token instead of waiting for the full response.
  • Vim-style navigation: j/k keys, number shortcuts (1-9), arrow keys.
  • Scriptable: --frontend=noninteractive and --output-format=json for automation. Pipe commands to shell-ai explain via stdin.
  • Configuration introspection: shell-ai config shows current settings and their sources.

Run shell-ai --help for all options, or shell-ai config schema for the full configuration reference.

Showcase

Suggest: XKCD #1168 (tar)

I don't know what's worse--the fact that after 15 years of using tar I still can't keep the flags straight, or that after 15 years of technological advancement I'm still mucking with tar flags that were 15 years old when I started.
shell-ai suggest -- any valid tar command to disarm the bomb

Explain: XKCD #1654 (Universal Install Script)

The failures usually don't hurt anything, and if it installs several versions, it increases the chance that one of them is right. (Note: The 'yes' command and '2>/dev/null' are recommended additions.)
printf '#!/bin/bash\n\npip install "$1" &\neasy_install "$1" &\nbrew install "$1" &\nnpm install "$1" &\nyum install "$1" & dnf install "$1" &\ndocker run "$1" &\npkg install "$1" &\napt-get install "$1" &\nsudo apt-get install "$1" &\nsteamcmd +app_update "$1" validate &\ngit clone https://github.com/"$1"/"$1" &\ncd "$1";./configure;make;make install &\ncurl "$1" | bash &' | shell-ai explain

Multilingual

Suggest in Danish (Foreslå på dansk) Explain in French (Expliquer en français)
shai Oversæt rødgrød med fløde til engelsk med Ollama API og model gemma3:27b-cloud shell-ai --locale fr_FR explain -- 'sudo !!'

Challenging Tasks

Suggest Explain
shell-ai suggest 'perl: Animate 5 seconds of an indeterminate progress bar in the style of a six-color ANSI rainbow wave at 24 frames per second' shell-ai explain -- rsync --delete-delay --delay-updates -avyyHXShPs

JSON Output for Scripting

shell-ai --frontend=noninteractive --output-format=json explain -- ls -lhtr | jq '.'

Configuration Introspection

SHAI_SKIP_CONFIRM=true shell-ai config

Configuration

Shell-AI loads configuration from multiple sources (highest priority first):

  1. CLI flags (--provider, --model, etc.)
  2. Environment variables (SHAI_API_PROVIDER, OPENAI_API_KEY, etc.)
  3. Config file (see paths below)
  4. Built-in defaults

Config file locations:

  • Linux: ~/.config/shell-ai/config.toml
  • macOS: ~/Library/Application Support/shell-ai/config.toml
  • Windows: %APPDATA%\shell-ai\config.toml

Generate a documented config template:

shell-ai config init

Example config:

provider = "openai"

[openai]
api_key = "sk-..."
model = "gpt-4o"

Providers

Set the provider in your config file:

  • Linux: ~/.config/shell-ai/config.toml
  • macOS: ~/Library/Application Support/shell-ai/config.toml
  • Windows: %APPDATA%\shell-ai\config.toml

The provider-specific settings go in a section named after the provider.

provider = "openai"  # or: anthropic, claudecode

Shell-AI may alternatively be configured by environment variables, which override the config file:

export SHAI_API_PROVIDER=openai  # or: anthropic, claudecode

[!TIP] Run shell-ai config schema to see all available settings and their defaults.

OpenAI-Compatible Providers

The providers openai, groq, ollama, and mistral all use the OpenAI chat completions API format. They share the same configuration structure and support temperature and max_tokens settings.

Provider Environment Variables
openai OPENAI_API_KEY, OPENAI_API_BASE, OPENAI_MODEL, OPENAI_MAX_TOKENS, OPENAI_ORGANIZATION
groq GROQ_API_BASE, GROQ_API_KEY, GROQ_MODEL, GROQ_MAX_TOKENS
ollama OLLAMA_API_BASE, OLLAMA_MODEL, OLLAMA_MAX_TOKENS
mistral MISTRAL_API_KEY, MISTRAL_API_BASE, MISTRAL_MODEL, MISTRAL_MAX_TOKENS

[!NOTE] provider = "openai" works with any OpenAI-compatible API.

The main differences from other OpenAI-compatible providers are the default API base URLs and models.

[openai]
api_key = "sk-..."
# api_base = "https://api.openai.com"  # change for compatible APIs
# model = ""
# max_tokens = ""
# organization = ""  # for multi-org accounts

export OPENAI_API_KEY=sk-...
# export OPENAI_API_BASE=https://api.openai.com
# export OPENAI_MODEL=
# export OPENAI_MAX_TOKENS=
# export OPENAI_ORGANIZATION=

[groq]
api_key = "gsk_..."
# api_base = "https://api.groq.com/openai"
# model = ""
# max_tokens = ""

export GROQ_API_KEY=gsk_...
# export GROQ_API_BASE=https://api.groq.com/openai
# export GROQ_MODEL=
# export GROQ_MAX_TOKENS=

[ollama]
# api_base = "http://localhost:11434"
# model = ""
# max_tokens = ""

# export OLLAMA_API_BASE=http://localhost:11434
# export OLLAMA_MODEL=
# export OLLAMA_MAX_TOKENS=

[mistral]
api_key = "your-key"
# api_base = "https://api.mistral.ai"
# model = ""
# max_tokens = ""

export MISTRAL_API_KEY=your-key
# export MISTRAL_API_BASE=https://api.mistral.ai
# export MISTRAL_MODEL=
# export MISTRAL_MAX_TOKENS=

Azure OpenAI

Azure OpenAI uses a different URL structure with deployment names instead of model selection.

[azure]
api_key = "your-key"  # REQUIRED
api_base = "https://your-resource.openai.azure.com"  # REQUIRED
deployment_name = "your-deployment"  # REQUIRED
# api_version = "2023-05-15"
# max_tokens = ""

export AZURE_API_KEY=your-key  # REQUIRED
export AZURE_API_BASE=https://your-resource.openai.azure.com  # REQUIRED
export AZURE_DEPLOYMENT_NAME=your-deployment  # REQUIRED
# export OPENAI_API_VERSION=2023-05-15
# export AZURE_MAX_TOKENS=

Anthropic

Use the native Anthropic Messages API as an alternative to the OpenAI chat completions API.

[!NOTE] provider = "anthropic" works with any Anthropic-compatible API, like Ollama v0.14.0.

[anthropic]
api_key = "sk-ant-..."
# api_base = "https://api.anthropic.com"
# model = ""
# max_tokens = ""

export ANTHROPIC_API_KEY=sk-ant-...
# export ANTHROPIC_API_BASE=https://api.anthropic.com
# export ANTHROPIC_MODEL=
# export ANTHROPIC_MAX_TOKENS=

Claude Code

Uses the Claude Code CLI in non-interactive mode. No API key or login configuration needed; Claude Code manages its own authentication.

[claudecode]
# cli_path = "claude"  # path to claude executable
# model = ""           # e.g., haiku, sonnet, opus

# export CLAUDE_CODE_CLI_PATH=claude
# export CLAUDE_CODE_MODEL=

Requirements: Claude Code CLI installed and authenticated (claude command available in PATH, or specify full path via cli_path).

Shell Integration

Shell-AI works well standalone, but integrating it into your shell enables any or all of these streamlined workflows:

  • Tab completion for shell-ai commands
  • Aliases as shorthands for shell-ai commands:
    • ?? alias for shell-ai suggest --
    • explain alias for shell-ai explain --
  • Ctrl+G keybinding to transform the current line into a shell command

Setup

Generate an integration file for your shell:

# Generate with default features (completions + aliases)
shell-ai integration generate bash

# Or with all features including Ctrl+G keybinding
shell-ai integration generate bash --preset full

Then add the source line to your shell config as instructed.

Available presets:

Feature minimal standard full
Tab completions ✓ ✓ ✓
Aliases (??, explain) ✓ ✓
Ctrl+G keybinding for suggest ✓

Default: standard

Customization examples:

# Standard preset plus keybinding
shell-ai integration generate zsh --preset standard --add keybinding

# Full preset without aliases
shell-ai integration generate fish --preset full --remove aliases

# Update all installed integrations after upgrading shell-ai
shell-ai integration update

# View available features and installed integrations
shell-ai integration list

Alternative: eval on startup (not recommended)

Instead of generating a static file, you can eval the integration directly in your shell config:

# Bash/Zsh
eval "$(shell-ai integration generate bash --preset=full --stdout)"
eval "$(shell-ai integration generate zsh --preset=full --stdout)"

# Fish
shell-ai integration generate fish --preset=full --stdout | source

# PowerShell
Invoke-Expression (shell-ai integration generate powershell --preset=full --stdout | Out-String)

This approach doesn't write files to your config directory and is always up to date after upgrading Shell-AI, but adds several milliseconds to shell startup (the time to spawn Shell-AI and generate the integration). The file-based approach above is recommended for faster startup.

Performance

The shell integration file is pre-compiled to minimize shell startup overhead. Here are benchmark results comparing the overhead of each preset.

This is how much slower Shell-AI v0.5.1's shell integration makes shell startup:

Baseline: Sourcing an Empty File

Shell N Min Q1 Median Q3 Max Mean Std Dev
Bash 1000 1.06ms 1.18ms 1.21ms 1.31ms 2.95ms 1.27ms 0.16ms
Zsh 1000 1.17ms 1.33ms 1.37ms 1.45ms 4.87ms 1.42ms 0.23ms
Fish 1000 0.78ms 0.88ms 0.91ms 0.96ms 2.69ms 0.94ms 0.12ms
PowerShell 100 79.03ms 81.09ms 82.48ms 84.91ms 98.50ms 83.32ms 3.31ms

Incremental Overhead (Above Baseline)

Shell Preset Overhead (Mean)
Bash minimal +1.56ms
Bash standard +1.64ms
Bash full +2.11ms
Zsh minimal +1.98ms
Zsh standard +2.05ms
Zsh full +2.43ms
Fish minimal +2.42ms
Fish standard +2.56ms
Fish full +2.69ms
PowerShell minimal +20.30ms
PowerShell standard +21.67ms
PowerShell full +125.24ms

Total Overhead (What Users Experience)

Bash
Preset N Min Q1 Median Q3 Max Mean Std Dev
blank (baseline) 1000 1.06ms 1.18ms 1.21ms 1.31ms 2.95ms 1.27ms 0.16ms
minimal 1000 2.55ms 2.71ms 2.78ms 2.89ms 3.85ms 2.82ms 0.17ms
standard 1000 2.62ms 2.76ms 2.85ms 2.98ms 6.08ms 2.91ms 0.25ms
full 1000 2.97ms 3.22ms 3.32ms 3.47ms 6.55ms 3.38ms 0.26ms
Zsh
Preset N Min Q1 Median Q3 Max Mean Std Dev
blank (baseline) 1000 1.17ms 1.33ms 1.37ms 1.45ms 4.87ms 1.42ms 0.23ms
minimal 1000 3.03ms 3.27ms 3.35ms 3.47ms 5.07ms 3.40ms 0.20ms
standard 1000 3.07ms 3.32ms 3.41ms 3.55ms 5.86ms 3.47ms 0.25ms
full 1000 3.40ms 3.69ms 3.80ms 3.94ms 6.20ms 3.85ms 0.27ms
Fish
Preset N Min Q1 Median Q3 Max Mean Std Dev
blank (baseline) 1000 0.78ms 0.88ms 0.91ms 0.96ms 2.69ms 0.94ms 0.12ms
minimal 1000 3.03ms 3.19ms 3.27ms 3.44ms 4.64ms 3.36ms 0.26ms
standard 1000 3.15ms 3.29ms 3.40ms 3.64ms 5.50ms 3.50ms 0.29ms
full 1000 3.30ms 3.45ms 3.53ms 3.70ms 5.58ms 3.63ms 0.27ms
PowerShell
Preset N Min Q1 Median Q3 Max Mean Std Dev
blank (baseline) 100 79.03ms 81.09ms 82.48ms 84.91ms 98.50ms 83.32ms 3.31ms
minimal 100 96.98ms 101.39ms 103.04ms 105.27ms 118.27ms 103.62ms 3.95ms
standard 100 99.77ms 103.34ms 104.47ms 106.13ms 115.67ms 104.98ms 3.12ms
full 100 200.65ms 205.15ms 207.09ms 209.94ms 241.00ms 208.55ms 5.94ms

Methodology

To reproduce these benchmarks, run cargo run --package xtask -- bench-integration [sample_count] from this repository.

Migrating from Python Shell-AI

If you're coming from ricklamers/shell-ai:

  • The provider is required. Set SHAI_API_PROVIDER explicitly, as the default is no longer Groq.
  • SHAI_SKIP_HISTORY is removed. Writing to shell history is no longer supported. The previous implementation made assumptions about the shell's history configuration. Shells don't expose history hooks to child processes, making this feature infeasible.
  • SHAI_SKIP_CONFIRM is deprecated. Use --frontend=noninteractive or SHAI_FRONTEND=noninteractive as a more flexible alternative.
  • Context mode is deprecated. The --ctx flag and CTX environment variable still work but are not recommended. The extra context from shell output tends to confuse the completion model rather than help it.
  • Model defaults differ. Set model explicitly if you prefer a specific model.

Contributing

Contributions welcome! Open an issue or pull request at Deltik/shell-ai.

For changes to the original Python Shell-AI, head upstream to ricklamers/shell-ai.

Acknowledgments

This project began as a fork of ricklamers/shell-ai at v0.4.4. Since v0.5.0, it shares no code with the original—a complete Ship of Theseus rebuild in Rust. The hull is new, but the spirit remains.

License

Shell-AI is licensed under the MIT License. See LICENSE for details.