sgpt

Converts natural language descriptions into executable shell commands by sending user intent to LLM APIs (OpenAI or compatible) and parsing structured command output. The tool maintains shell context awareness, allowing it to generate commands tailored to the user's current environment and shell type (bash, zsh, fish, etc.). Output is presented for user review before execution, with optional one-shot execution mode for trusted workflows.

Provides a multi-turn conversational interface within the terminal where users can ask follow-up questions and refine LLM responses iteratively. The tool maintains conversation history across turns, allowing context carryover for related queries. Chat mode operates as a REPL-like loop, accepting user input, sending to the LLM with full conversation context, and streaming responses back to the terminal with proper formatting.

Maintains conversation state across multiple turns in chat mode, preserving full message history and context for the LLM. Each turn includes the user's new message plus all previous messages, allowing the LLM to reference earlier parts of the conversation. State is held in memory during the session and can be optionally exported or saved to files for later retrieval.

Generates code snippets in multiple programming languages (Python, JavaScript, Go, Rust, etc.) from natural language descriptions. The tool sends language-specific prompts to the LLM and returns formatted code blocks suitable for copy-paste or piping to files. Code generation respects language context when available (e.g., if invoked from a Python project, defaults to Python output).

Integrates sgpt output directly into shell pipelines and command substitution contexts, allowing LLM-generated content to feed into other commands or be stored in variables. The tool outputs plain text suitable for shell consumption, enabling patterns like `$(sgpt 'generate a JSON config')` or `sgpt 'list files' | grep pattern`. Integration respects shell quoting and escaping conventions to prevent injection vulnerabilities.

Abstracts LLM API interactions to support OpenAI and compatible endpoints (e.g., Azure OpenAI, local Ollama instances, or other OpenAI-compatible APIs). Configuration is managed via environment variables or config files, allowing users to switch providers without code changes. The tool handles API authentication, request formatting, and response parsing transparently across providers.

Constructs LLM prompts with system instructions and context that tailor responses to specific use cases (shell commands, code generation, explanations, etc.). The tool embeds domain-specific prompting strategies that guide the LLM toward generating safe, executable, and relevant output. System prompts are customizable via configuration, allowing users to inject project-specific guidelines or constraints.

Streams LLM responses token-by-token to the terminal as they arrive, rather than buffering the entire response before display. This provides real-time feedback and reduces perceived latency for long responses. The tool handles terminal rendering, line wrapping, and ANSI color codes to present streamed output cleanly. Streaming is compatible with piping and command substitution, though buffering may occur in those contexts.

Integrates with shell history mechanisms (bash history, zsh history, etc.) to track generated commands and enable recall of previous LLM outputs. The tool can optionally cache frequently-used command patterns to reduce API calls and latency. Caching is transparent to the user and respects cache invalidation policies based on time or explicit user action.

Implements safety mechanisms to prevent execution of potentially dangerous shell commands, including command review before execution, optional dry-run mode, and warnings for risky patterns (e.g., `rm -rf`, `sudo` commands). The tool provides a confirmation prompt by default, allowing users to review and edit commands before execution. Advanced users can disable guardrails via configuration for trusted workflows.

Manages tool configuration through environment variables (e.g., OPENAI_API_KEY, SGPT_MODEL) and optional YAML/TOML config files, allowing users to customize behavior without CLI flags. Configuration is hierarchical: environment variables override config files, which override defaults. This enables both global system-wide configuration and per-project overrides.