v2.1.1 Now Available

Create any app you want with AI in your terminal

A high-performance, terminal-based AI pair programmer. Combining state-of-the-art LLMs with a rich TUI and robust tool-calling to help you write, debug, and explore code without leaving your CLI.

Install globally via npm:

npm install -g @banaxi/banana-code

⚠️ Ensure you install @banaxi/banana-code.

🤔 Why Banana Code?

Built for developers who live in the terminal and want maximum flexibility.

No Vendor Lock-in

Switch instantly between proprietary models (Gemini, Claude, OpenAI) and open-source models (Ollama Local, Ollama Cloud, Mistral) mid-conversation.

True Autonomy

With Plan & Execute mode and Self-Healing Error Loops, Banana Code tries, fails, reads errors, and fixes its own mistakes automatically.

Terminal Native

Brings workspace awareness, web search, and surgical file patching directly to your CLI without forcing you to change your IDE.

✨ Key Features

Multi-Provider Support

Supports Google Gemini, Anthropic Claude, OpenAI, Mistral AI, OpenRouter, and Ollama. Seamlessly switch whenever you need.

Auto Mode

A smart router model reads your prompt and automatically selects the best model and reasoning effort for that specific turn.

Interactive Terminal

AI can spawn persistent terminal sessions to handle interactive prompts (like npm init or Y/N confirmations) in real-time.

Financial Intelligence

Track exact API spend and savings. Shows session costs and money saved via Prompt Caching.

MCP Support

Connect to community-built Model Context Protocol servers (SQLite, GitHub, Google Maps) to give your AI infinite new superpowers.

Banana Guard

AI-Powered Auto-Approve. Automatically approves safe tasks while stopping risky shell operations for your review. YOLO mode, safely.

Self-Healing Loop

If a command fails, Banana Code automatically feeds the error trace back to the AI so it can fix its own code without your help.

Agent Skills

Teach your AI workflows. Drop a SKILL.md in your config folder, and the AI will auto-activate it when relevant.

Web Research

Deep integration with DuckDuckGo APIs and Scrapers gives the AI real-time access to the internet to find docs and solutions.

🆚 How It Compares

See why Banana Code stands out against other popular terminal tools.

Feature 🍌 Banana Code Claude Code Gemini CLI
Multi-Model Support (OpenRouter/Mistral) ✅ ❌ ❌
Local Privacy (Ollama / Local LLMs) ✅ ❌ ❌
Custom Agent Skills (SKILL.md) ✅ Partial ✅
Interactive Stateful Terminal ✅ Partial ❌
Auto-Mode Routing ✅ ❌ Partial
Model Context Protocol (MCP) ✅ ✅ ❌
Global AI Memory & Preferences ✅ ✅ ✅
Transparent Cost Tracking ✅ ✅ ❌
Open Source Engine ✅ ❌ ✅
And a lot more... ✅ - -

🛠️ Installation & Setup

Get started in seconds. Install globally, run the tool, and follow the simple setup wizard.

1. Install

npm install -g @banaxi/banana-code

2. Run Setup

banana

3. Resume Later (Optional)

banana --resume

Welcome to Banana Code Setup!

Please select your preferred provider:

❯ Google Gemini

Anthropic Claude

OpenAI

Ollama (Local)

Enter API Key:

************************

Setup complete! Type a message or /help.

You: build a tic tac toe app

In-App Commands

While in a chat, use these special commands to control the environment.

Command Description
/provider Switch provider: gemini, claude, openai, mistral, openrouter, ollama_cloud, ollama
/model Change model; omit the name to open the menu (includes Auto Mode where supported).
/chats Browse and resume saved sessions (auto-titled).
/clean Summarize long history into a short memory to save tokens.
/context Show message count, estimated tokens, and financial spend/savings.
/beta Access Beta tools like MCP and optional scrapers.
/init Generate a BANANA.md project summary in the current directory.
@file/path.js Instantly inject file contents into your prompt context.

🎛️ Operating Modes

Layer specific behavioral modes on top of the agent for safe and targeted execution.

/agent (Default)

Full coding agent with tools. It applies changes directly to your filesystem subject to your permissions.

/plan

AI proposes a written plan before large edits.

  • Small changes applied directly.
  • Significant work requires explicit plan approval.

/ask

Read-only Q&A mode.

  • No file modifications allowed.
  • Can run read-only tools (git status).

/security

Prioritize finding vulnerabilities.

  • Focus on OWASP-style issues.
  • Provides actionable remediation ideas.

/skill-creator

Expert prompt engineer mode.

  • Generates custom Agent Skills automatically.
  • Saves directly to config folder.

/yolo

Auto-approve all permissions.

  • Bypasses standard manual approvals.
  • Dangerous; use with extreme care.

Agent Skills

Skills act as "onboarding guides" teaching the AI specific tasks or company standards. The AI auto-detects tasks and activates the relevant skill.

Create a skill at ~/.config/banana-code/skills/my-react-skill/SKILL.md

---
name: my-react-skill
description: Use when editing a React component.
---
# React Guidelines
- Always use functional components.
- Always use Tailwind CSS for styling.
- Do not use default exports.

Model Context Protocol (MCP)

Plug in community-built servers to give your AI access to your databases, GitHub, Slack, and more via the MCP standard.

  • Enable MCP Support in the /beta menu.
  • Create a config file at ~/.config/banana-code/mcp.json.
  • Restart Banana Code to auto-load the tools natively!
{
  "mcpServers": {
    "everything": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-everything"]
    }
  }
}

Headless API Mode

Run Banana Code as a background engine exposing its tool-calling logic via HTTP and WebSockets. Build custom GUIs on top of our engine!

banana --api 4000

Protected by Secure API Tokens generated at startup.

// WebSocket Connect (ws://localhost:4000?token=...)
// Send chat:
{ "type": "chat", "text": "Run tests" }

// Remote Tool Approval Ticket
{
  "type": "permission_requested",
  "ticketId": "5c9b2a...",
  "action": "Execute Command"
}