What is miniclaude?

MiniClaude is a lightweight terminal based AI coding assistant modeled after Claude Code. Instead of running as a single throwaway command line process, it is built as a local first system with a background daemon called mini-core that stays running and handles sessions, tool calls, events, permission approvals, and calls to the Anthropic API. Two separate clients, a plain command line tool called mini and a full terminal interface called mini-tui, talk to this daemon over a local network connection and let a person chat with the assistant or run one off tasks. Because the daemon and the clients are separate, other tools could in theory connect to the same daemon and watch or control what the assistant is doing, since everything happens through a documented JSON based protocol rather than being locked inside one process. The terminal interface shows the conversation, streaming responses, tool calls, and permission prompts in real time, and can even replay a past run from its saved event log. Built in tools let the assistant read and write files, run shell commands, manage simple tasks and notes, and delegate work to sub agents. Every tool call passes through a permission system, where a user can allow or deny a specific action either once or permanently, with the choices saved to a local policy file. The assistant also supports the Model Context Protocol, so external MCP tool servers can be plugged in and their tools become available to the assistant automatically. Other features mentioned include saving context and notes across turns, compacting long conversation history so it does not grow without bound, and a trace log that records internal events for debugging and replay. The project is written in Python, managed with the uv package manager, and requires Python 3.12 along with an Anthropic API key to actually run a chat or task. It is released under the MIT license.