← Back to Agent ⬡
Repository nano-coder
↗

Agent

Nano Coder

Minimal terminal-based coding agent with tool use, streaming answers, MCP support, and local subagents for code execution and development workflows.

Why This Is Important for AGI Infrastructure

Agents are increasingly central operators of software systems. AGI infrastructure needs agent tooling that is transparent, controlled, and grounded in real system constraints.

Specifically for Nano Coder, this work creates a clearer path from system design decisions to measurable AGI workload behavior.

What This Is About

Minimal terminal-based coding agent with tool use, streaming answers, MCP support, and local subagents for code execution and development workflows.

Overview

nano-coder is a minimal terminal-based coding agent built in feature-focused steps from a minimal agent to a capable repo-workbench. It demonstrates how to build a practical tool-use agent system with streaming answers, MCP (Model Context Protocol) support, local subagents, and a skill system.

What Nano-Coder Is

  • A terminal-first coding agent that helps with code tasks through tool use
  • A streaming answer system that provides real-time feedback
  • An extensible skill system for adding new capabilities
  • A planning workflow with /plan command for complex tasks
  • A reference implementation for building minimal but capable agents

Key Features

  • Terminal-first workflow: Designed for command-line usage
  • Tool-based code tasks: Read files, edit code, run commands
  • Streaming answers: Real-time output as the agent thinks
  • MCP support: Model Context Protocol for extended capabilities
  • Local subagents: Specialized agents for specific tasks
  • Skill system: Extensible command framework
  • Context compaction: Efficient context management
  • Per-session logging: Track agent behavior
  • Slash commands: Quick access to common operations

Architecture

Built in feature-focused steps:

  1. Minimal agent: Basic tool use and streaming
  2. Planning workflow: /plan command for complex multi-step tasks
  3. Skill system: Extensible command framework
  4. MCP integration: Model Context Protocol support
  5. Local subagents: Specialized helper agents

Development Journey

  • Started as a minimal terminal agent with basic tool use
  • Added streaming answers for real-time feedback
  • Implemented /plan workflow for complex tasks
  • Built extensible skill system
  • Added MCP support for extended capabilities
  • Created local subagents for specialized tasks

This project demonstrates how to build a practical coding agent from first principles, focusing on minimal but functional implementations rather than complex frameworks.