MiniCode: Lightweight Terminal Coding Assistant with Claude Code Workflow

What changed

MiniCode has been introduced as a lightweight, terminal-based coding assistant. It aims to replicate the workflow and architectural concepts of Claude Code in a significantly smaller implementation, making it suitable for learning and experimentation. The project is built around a terminal-first agent loop that accepts user requests, inspects the workspace, calls tools when necessary, reviews file changes before committing them, and provides a final response within the same terminal session. The codebase is intentionally compact to facilitate understanding and extension of its control flow, tool model, and TUI behavior.

Core Capabilities

• Multi-step tool execution: Supports a model -> tool -> model loop within a single turn.
• Concurrent sub-agents: Up to three read-only sub-agents can operate concurrently, with the root agent managing code changes.
• In-memory planning: Features a shared Todo plan across agent turns, viewable with /plan.
• Persistent Goals: Allows for user-created Goals that persist across turns, with options to pause, resume, and explicitly complete.
• Recurring prompts: A process-local loop can repeat prompts at fixed intervals, sharing the Plan and remaining mutually exclusive with active Goals.
• Full-screen TUI: Includes input history, transcript scrolling, a slash command menu, and approval flows.
• Session persistence: Supports per-project session saving with commands for resume, rename, fork, and compact.
• Context management: Implements provider-usage-first context statistics with tail estimates, auto-compact, context collapse, and snip compact.
• Built-in tools: Offers tools for file operations, search, editing, command execution, web fetching, and clarification prompts.
• Local skills and MCP: Discovers local skills via SKILL.md and integrates with MCP tools/resources/prompts over stdio or remote HTTP.
• Review-before-write: Edits are presented for review with path and command permission checks.
• Oversized result handling: Large tool outputs are stored on disk and represented by a short preview and file path in the context.

#

Why it matters for builders

MiniCode provides a valuable resource for developers looking to understand and experiment with AI agent architectures without the complexity of larger platforms. Its modular design and focus on a clear control flow make it an excellent educational tool. Builders can use it as a reference implementation for Claude Code-like agent patterns or adapt its components for their own custom terminal-based AI tooling.

#

Practical impact

Developers can install MiniCode by cloning the repository and running npm install and npm run install-local. The installation process requires specifying the model name, Anthropic base URL, and authentication token. Configuration is stored in ~/.mini-code/settings.json and ~/.mini-code/mcp.json. A quick start involves running the installed launcher with the minicode command. For development, npm run dev can be used, and an offline demo mode is available via MINI_CODE_MODEL_MODE=mock npm run dev. The project also offers multiple language implementations, including Rust (MiniCode-rs), Python (MiniCode-Python), Go (MiniCode-go), and Java (MiniCode4j), alongside the primary TypeScript version.

#

Caveats and source limits

The source provides detailed information about MiniCode's architecture, capabilities, and installation. However, specific performance benchmarks, pricing details for any associated services, or precise release dates beyond the general availability of the project are not included. The project is presented as a tool for learning and experimentation, and its current stage of development or production readiness is not explicitly stated. The provided star and fork counts (748 stars, 84 forks) are static and do not reflect real-time activity.