👾 mini-coder

Small. Fast. Gets out of your way.

Hey there! I'm mini-coder — a CLI coding agent built for developers who want a sharp tool, not a bloated IDE plugin. Think of me as the pocket knife of AI coding assistants: lightweight, reliable, and always ready.

🤙 Who Am I?

I'm mc — your new terminal companion. I live in your shell, speak to large language models, and help you explore, understand, and modify code at the speed of thought.

I was built with a simple philosophy: dev flow first. No slow startup. No clunky GUI. No vendor lock-in. Just you, your terminal, and an AI that keeps up.

$ mc
┌─ mini-coder ──────────────────────────────────────────┐
│  What would you like to work on today?                │
│                                                       │
│  > _                                                  │
│                                                       │
│  [zen/claude-sonnet-4-6] [~/src/my-project] [main] ...│
└───────────────────────────────────────────────────────┘

🛠️ What Can I Do?

I come equipped with a tight, reliable set of tools:

| Tool | What it does | |---|---| | 🔍 glob | Find files by pattern across your project | | 🧲 grep | Search file contents with regex | | 📖 read | Read files (with line-range support) | | 📝 create | Create a new file or fully overwrite an existing file | | ✏️ replace | Replace or delete lines using hashline anchors | | ➕ insert | Insert lines before/after an anchor without replacing | | 🐚 shell | Run shell commands and see their output | | 🤖 subagent | Spawn a focused mini-me for parallel subtasks |

I can also connect to MCP servers (like Exa for web search), giving you superpowers on demand.

⚡ Key Features

• Multi-provider — set OPENCODE_API_KEY for Zen, ANTHROPIC_API_KEY, OPENAI_API_KEY, GOOGLE_API_KEY, or just run Ollama locally. I auto-discover whatever's available.
• Session memory — conversations are saved in a local SQLite database. Resume where you left off with -c or pick a specific session with -r <id>.
• Shell integration — prefix with ! to run shell commands inline. Use @ to reference files in your prompt (with Tab completion).
• Slash commands — /model to switch models, /plan for read-only thinking mode, /ralph for autonomous looping, /review for a code review, /undo to roll back a turn, /new for a clean session, /mcp to manage MCP servers. See all with /help.
• Custom commands — drop a .md file in .agents/commands/ and it becomes a /command. Supports argument placeholders ($ARGUMENTS, $1…$9) and shell interpolation (!`cmd`). Global commands live in ~/.agents/commands/. Custom commands take precedence over built-ins. → docs/custom-commands.md
• Custom agents — drop a .md file in .agents/agents/ (or ~/.agents/agents/ globally) and reference it with @agent-name in your prompt. The agent runs in its own context window with a custom system prompt and optional model override. → docs/custom-agents.md
• Skills — place a SKILL.md in .agents/skills/<name>/ and inject it into any prompt with @skill-name. Skills are never auto-loaded — always explicit. → docs/skills.md
• Post-tool hooks — drop an executable at .agents/hooks/post-<tool> and I'll run it after every matching tool call.
• Beautiful, minimal output — diffs for edits, formatted trees for file searches, a live status bar with model, git branch, and token counts.
• 16 ANSI colors only — my output inherits your terminal theme. Dark mode, light mode, Solarized, Gruvbox — I fit right in.

🧠 Interesting Things About Me

• I eat my own dog food. I was built by a mini-coder agent. It's agents all the way down. 🐢
• I'm tiny but mighty. The whole runtime is Bun.js — fast startup, native TypeScript, and a built-in SQLite driver.
• I respect existing conventions. Hook scripts live in .agents/hooks/, context in AGENTS.md or CLAUDE.md, commands in .agents/commands/, agents in .agents/agents/, skills in .agents/skills/ — I follow the ecosystem instead of inventing new standards.
• I spin while I think. ⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏ (It's the little things.)
• I can clone myself. The subagent tool lets me spin up parallel instances of myself to tackle independent subtasks simultaneously. Divide and conquer! (Up to 3 levels deep.)

📁 The .agents folder

mini-coder follows the .agents convention used across the AI coding tool ecosystem. Drop files in .agents/ to extend behaviour for the current repo, or ~/.agents/ to apply them globally.

| Path | What it does | |---|---| | .agents/commands/*.md | Custom slash commands (/name) | | .agents/agents/*.md | Custom agents (@name) | | .agents/skills/<name>/SKILL.md | Reusable skill instructions (@name) | | .agents/hooks/post-<tool> | Scripts run after a tool call | | AGENTS.md | Project context injected into every system prompt |

Local always overrides global. The same ~/.agents/ folder is shared with Claude Code, Opencode, and other compatible tools — skills and agents you write once work everywhere.

🚀 Getting Started

# Install from npm
bun add -g mini-coder
# or: npm install -g mini-coder

# Set your provider key (pick one — or run Ollama locally)
export OPENCODE_API_KEY=your-zen-key    # recommended
export ANTHROPIC_API_KEY=your-key       # or direct Anthropic
export OPENAI_API_KEY=your-key          # or direct OpenAI

# Launch!
mc

Or run directly for a quick task:

mc "Refactor the auth module to use async/await"

Useful flags:

mc -c                        # continue last session
mc -r <id>                   # resume a specific session
mc -l                        # list recent sessions
mc -m zen/claude-sonnet-4-6  # pick a model

🗂️ Project Structure

src/
  index.ts          # Entry point + CLI arg parsing
  agent/            # Main REPL loop + tool registry
  cli/              # Input, output, slash commands, markdown rendering
  llm-api/          # Provider factory + streaming turn logic
  tools/            # glob, grep, read, create, replace, insert, shell, subagent
                    #   + hashline anchors, diffs, hooks, snapshots
  mcp/              # MCP server connections
  session/          # SQLite-backed session & history management

🔮 Tech Stack

• Runtime: Bun.js — fast, modern, all-in-one
• LLM routing: AI SDK — multi-provider with streaming
• Colors: yoctocolors — tiny and terminal-theme-aware
• Schema validation: Zod
• Linting/formatting: Biome
• Storage: bun:sqlite — zero-dependency local sessions

💬 Philosophy

Accurate. Fast. Focused on the conversation.

I believe the best tools disappear into your workflow. I don't want to be the star of the show — I want you to ship great code, faster.

💬 What People Are Saying

"sean this is fucking sick" — vpr99 (eric)

Built with ❤️ and a healthy obsession with terminal aesthetics.