Plain English to shell commands, powered by Claude or Ollama.

pip install shellsage-ai

After installing, run the one-time setup:

shellsage init

# Find all large files in your home directory
shellsage "find files larger than 100MB in my home directory"

# Create a new git branch and switch to it
shellsage "create a new git branch called feature/login and switch to it"

# Compress a folder
shellsage "compress the folder ./logs into a tar.gz archive"

# Kill a process on port 8080
shellsage "kill whatever process is running on port 8080"

# Bulk rename all .jpeg files to .jpg
shellsage "rename all .jpeg files in this folder to .jpg"

shellsage chat opens an interactive REPL where the full conversation — your requests, the AI's responses, and actual command output — is remembered across turns. Use it when you want to refine commands, build on previous results, or have a back-and-forth with the AI.

shellsage chat

shellsage> list python files in this repo
  → ls **/*.py          [safe]   Run? y
  output: shellsage/agent.py, shellsage/chat.py ...

shellsage> now show only the ones modified in the last 7 days
  → find . -name "*.py" -mtime -7    [safe]   Run? y

shellsage> what was the output of the first command?
  The first command listed: shellsage/agent.py, shellsage/chat.py ...

shellsage> exit

Each turn automatically feeds the command output back to the AI, so follow-up requests are always context-aware.

shellsage chat --dry-run          # show commands, never execute
shellsage chat --explain          # explain each command before prompting
shellsage chat --provider claude  # override provider for this session

ShellSage can run entirely offline using Ollama.

ollama pull llama3.2        # recommended — fast and capable
ollama pull qwen2.5:3b      # lightweight, great for slower machines
ollama pull mistral         # most capable open model

ollama serve
# macOS: starts automatically from the menu bar app

curl http://localhost:11434/api/tags

Then run shellsage init and choose Ollama.

Usage: shellsage [INTENT] [OPTIONS]
       shellsage COMMAND [OPTIONS]

Single-shot mode:
  shellsage "your intent"   Translate plain English into shell commands.

Options (single-shot):
  --dry-run           Show commands without executing them.
  --explain           Show per-token breakdown before prompting.
  --provider TEXT     Override configured provider: claude or ollama.
  --help              Show this message and exit.

Commands:
  chat      Start an interactive multi-turn chat session.
  init      Run the first-time setup wizard.
  config    Show current config / re-run the setup wizard.
  history   Print command history. Use --clear to wipe it.

# Preview without running
shellsage "delete all .log files" --dry-run

# Get a detailed explanation of each command
shellsage "set up a Python virtual environment" --explain

# Override provider for this run
shellsage "list docker containers" --provider claude

# Start interactive chat mode
shellsage chat
shellsage chat --dry-run --provider ollama

# View saved command history
shellsage history

# Clear history
shellsage history --clear

# Reconfigure at any time
shellsage config

Choose how ShellSage handles command execution during shellsage init or shellsage config:

The mode applies to both single-shot and chat sessions.

ShellSage has a two-layer safety system:

The following commands are always blocked and can never run, regardless of context:

• rm -rf /, rm -rf /*, rm -rf ~
• dd if=/dev/zero, dd if=/dev/random, dd if=/dev/urandom
• mkfs, chmod -R 777 /
• Fork bombs: :(){ :|:& };:
• Piping untrusted URLs into a shell: curl | bash, wget -O- | sh

Each proposed command is shown with a color-coded panel:

You always see the command and its danger level before being asked to run it. Press e at the confirmation prompt for a detailed token-by-token explanation.

shellsage config

This re-runs the interactive setup wizard where you can switch between Claude and Ollama, change models, update your API key, and adjust the execution mode.

1. ShellSage detects your OS, shell, working directory, and installed tools.
2. It sends your English intent to the configured LLM with a structured prompt.
3. The LLM responds with a JSON plan containing one or more shell commands.
4. ShellSage validates each command against the safety blocklist.
5. You confirm (or explain, or skip) each command before it runs.
6. If a command fails, ShellSage automatically attempts to self-correct using the error output.
7. In chat mode, every exchange (including command output) is fed back into the conversation so follow-up requests are context-aware.

• New command: shellsage chat — interactive REPL that accumulates the full conversation history (requests, AI responses, command output) across turns.
• Self-correction in chat mode uses the full conversation context, not just the last error.
• Blocked commands in chat mode no longer exit the session — they are reported and the REPL continues.
• Both Claude and Ollama providers now accept a messages array for multi-turn conversations.

• New execution mode: auto_safe — safe commands run automatically; caution and destructive commands still prompt for confirmation.
• Execution mode is configured during shellsage init / shellsage config and persists to ~/.shellsage/config.toml.
• Default mode remains ask_all (prompt before every command) for safety.

• Persistent history — shellsage history now shows commands across sessions (stored in ~/.shellsage/history.json, capped at 200 entries).
• shellsage history --clear — wipe all saved history.
• Fixed routing bug where shellsage history and shellsage config were incorrectly forwarded to the LLM as intents instead of being handled as subcommands.
• Config command now displays current settings in a formatted table before offering to re-run the wizard.

• Single-shot mode: translate plain English to shell commands via Claude or Ollama.
• Two-layer safety system: hard blocklist + LLM-assigned danger levels (safe / caution / destructive).
• --dry-run flag to preview commands without executing.
• --explain flag for per-token command breakdown.
• --provider flag to override the configured provider per run.
• Automatic self-correction on command failure using the stderr output.
• First-time setup wizard (shellsage init).
• Context-aware prompts: OS, shell, working directory, and installed tools are injected into every request.

1. Fork the repository.
2. Create a feature branch: git checkout -b feature/my-feature
3. Install dev dependencies: pip install -e ".[dev]" (package name on PyPI: shellsage-ai)
4. Run tests: pytest tests/
5. Submit a pull request.

Please keep pull requests focused on a single concern and include tests for new functionality.

MIT — see LICENSE.