sgpt

Converts natural language descriptions into executable shell commands by sending user intent to LLM APIs (OpenAI, compatible endpoints) and parsing structured responses. The tool maintains shell context awareness, allowing it to generate commands appropriate for the user's current shell environment (bash, zsh, fish, etc.) and operating system. Responses are validated before execution to prevent dangerous operations.

Provides a persistent REPL-style chat interface where users can ask multi-turn questions about shell operations, code, and system tasks. Each exchange maintains conversation history sent to the LLM, enabling contextual follow-up questions. Generated shell commands can be executed directly from the chat interface with output captured and fed back into the conversation for iterative refinement.

Generates code snippets in multiple programming languages (Python, JavaScript, Go, etc.) from natural language specifications. The tool sends language hints and code context to the LLM and returns formatted, executable code. Supports inline code generation within shell workflows and standalone code file creation.

Abstracts LLM provider selection through configuration, supporting OpenAI's API and any compatible endpoint (local Ollama, Hugging Face, custom servers). Configuration is stored in environment variables or config files, allowing users to switch providers without code changes. The tool handles authentication, request formatting, and response parsing for different provider APIs.

Integrates with shell environments (bash, zsh, fish, PowerShell) to capture generated commands and execute them directly within the user's shell context. The tool can be invoked as a shell function or alias, allowing generated commands to access the user's environment variables, working directory, and shell history. Execution results are captured and optionally fed back into the chat interface.

Allows users to define custom prompt templates that inject context (shell type, OS, project information) into LLM requests. Templates can include placeholders for environment variables, file contents, and system information. This enables consistent, context-aware prompts without manual context specification on each invocation.

Maintains conversation history across multiple turns, sending the full chat context to the LLM with each request. This enables the LLM to understand follow-up questions, reference previous commands, and provide coherent multi-step guidance. Context is managed in memory during a session and can be optionally saved to disk for later retrieval.

Formats generated commands and code with syntax highlighting for terminal display, making output more readable and visually distinguishable from regular shell output. Supports multiple output formats (plain text, colored terminal output, markdown) and can optionally wrap output in code blocks or shell-specific formatting.

Manages sgpt configuration through a combination of environment variables and configuration files (typically ~/.config/sgpt/config.json or equivalent). Users can set API keys, model names, temperature, max tokens, and other LLM parameters through either mechanism, with environment variables taking precedence over config files.