A lightweight terminal coding assistant with Claude Code-like workflow, tool loop, and TUI architecture, built for learning and experimentation.
# Add to your Claude Code skills
git clone https://github.com/LiuMengxuan04/MiniCode简体中文 | Architecture | Learn Claude Code Design Through MiniCode | License
A lightweight terminal coding assistant for local development workflows.
MiniCode provides Claude Code-like workflow and architectural ideas in a much smaller implementation, making it especially useful for learning, experimentation, and custom tooling.
MiniCode is built around a practical terminal-first agent loop:
The project is intentionally compact, so the control flow, tool model, and TUI behavior remain easy to understand and extend.
MiniCode is a good fit if you want:
No comments yet. Be the first to share your thoughts!
SKILL.mdlist_filesgrep_filesread_filewrite_fileedit_filepatch_filemodify_filerun_commandload_skilllist_mcp_resourcesread_mcp_resourcelist_mcp_promptsget_mcp_promptgit diff)Ctrl+O expand/collapse plus wheel/page scrollingcd mini-code
npm install
npm run install-local
The installer will ask for:
ANTHROPIC_BASE_URLANTHROPIC_AUTH_TOKENConfiguration is stored in:
~/.mini-code/settings.json~/.mini-code/mcp.jsonThe launcher is installed to:
~/.local/bin/minicodeIf ~/.local/bin is not already on your PATH, add:
export PATH="$HOME/.local/bin:$PATH"
Run the installed launcher:
minicode
Run in development mode:
npm run dev
Run in offline demo mode:
MINI_CODE_MODEL_MODE=mock npm run dev
minicode mcp listminicode mcp add <name> [--project] [--protocol <mode>] [--env KEY=VALUE ...] -- <command> [args...]minicode mcp remove <name> [--project]minicode skills listminicode skills add <path> [--name <name>] [--project]minicode skills remove <name> [--project]/help/tools/skills/mcp/status/model/model <name>/config-pathsExample configuration:
{
"model": "your-model-name",
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
}
},
"env": {
"ANTHROPIC_BASE_URL": "https://api.anthropic.com",
"ANTHROPIC_AUTH_TOKEN": "your-token",
"ANTHROPIC_MODEL": "your-model-name"
}
}
Project-scoped MCP config is also supported through Claude Code compatible .mcp.json:
{
"mcpServers": {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "."]
}
}
}
For vendor compatibility, MiniCode now auto-negotiates stdio framing:
Content-Length framing is tried first"protocol": "content-length" or "protocol": "newline-json"Skills are discovered from:
./.mini-code/skills/<skill-name>/SKILL.md~/.mini-code/skills/<skill-name>/SKILL.md./.claude/skills/<skill-name>/SKILL.md~/.claude/skills/<skill-name>/SKILL.mdConfiguration priority:
~/.mini-code/settings.json~/.mini-code/mcp.json.mcp.jsonMiniCode supports two extension layers:
skills: local workflow instructions, usually described by a SKILL.mdMCP: external tool providers that expose tools, resources, and prompts into MiniCodeInstall a local skill:
minicode skills add ~/minimax-skills/skills/frontend-dev --name frontend-dev
List installed or discovered skills:
minicode skills list
Inside the interactive UI, you can also run:
/skills
to inspect which skills are available in the current session.
If you explicitly mention a skill name, MiniCode will prefer loading it. For example:
Use the frontend-dev skill and directly rebuild the current landing page instead of stopping at a plan.
If you want to be even more explicit:
Load the fullstack-dev skill first, then follow its workflow to implement this task.
A common pattern is to clone an official or Claude Code-compatible skills repo locally and install from there:
git clone https://github.com/MiniMax-AI/skills.git ~/minimax-skills
minicode skills add ~/minimax-skills/skills/frontend-dev --name frontend-dev
Install a user-scoped MCP server:
minicode mcp add MiniMax --env MINIMAX_API_KEY=your-key --env MINIMAX_API_HOST=https://api.minimaxi.com -- uvx minimax-coding-plan-mcp -y
List configured MCP servers:
minicode mcp list
To configure an MCP server only for the current project, add --project:
minicode mcp add filesystem --project -- npx -y @modelcontextprotocol/server-filesystem .
minicode mcp list --project
Inside the interactive UI, run:
/mcp
to see which servers are connected, which protocol they negotiated, and how many tools / resources / prompts they expose.
MCP tools are automatically registered as:
mcp__<server_name>__<tool_name>
For example, after connecting the MiniMax MCP server you may see:
mcp__minimax__web_searchmcp__minimax__understand_imageThese tool names are not hand-written in MiniCode. They appear automatically after a successful MCP connection.
The simplest approach is to just describe the task naturally and let the model decide when to use a skill or MCP tool:
Search for recent Chinese-language resources about MCP and give me 5 representative links.
If MiniMax MCP is connected, the model will typically choose mcp__minimax__web_search.
If you want a more controlled workflow, name the skill or target capability explicitly:
Use the frontend-dev skill and directly modify the current project files to turn this page into a more complete product landing page.
Or:
Use the connected MCP tools to search for the MiniMax MCP guide and summarize what capabilities it provides.
skills are better for workflow, conventions, domain-specific instructions, and reusable execution patternsMCP is better for search, image understanding, browsers, filesystems, databases, and other remote capabilitiesA common combination is:
frontend-dev to shape how the work should be doneMiniCode currently focuses on:
SKILL.md discovery with load_skillFor vendor compatibility, MiniCode automatically tries:
Content-Length framingnewline-json if neededThat means servers such as MiniMax MCP, which use newline-delimited JSON over stdio, can still be connected directly.
If you want to study the project as a learning resource, continue with: