mramma 🚀

A powerful, multi-provider autonomous coding agent for your terminal.

mramma is a highly capable, terminal-based AI coding assistant designed to give you autonomous engineering capabilities right where you work. It can read your codebase, write files, search directories, and execute shell commands to accomplish complex tasks efficiently.

Features ✨

• 🧠 Multi-Provider Support: Seamlessly use Anthropic (Claude), Google Gemini, GitHub Models, OpenRouter, or Nvidia APIs.
• ⚡ Multiple Agent Modes: Choose between full Agent autonomy, Read-only Chat, Fast responses, or Markdown-based Planning.
• 🛠️ Built-In Tooling: Built-in capabilities to view files, list directories, search with grep, write/edit code, and run shell commands.
• 💰 Token Economics: Live tracking of token usage and session cost estimation.
• 🔌 Highly Extensible: Supports customized system prompts via project hooks and intelligent prompt execution tiers.
• 🔄 Session Resumption: Easily pick up where you left off with persistent local sessions.
• ⏱️ Live Reasoning Streams: Built-in support for deep thinking models (like DeepSeek and GLM), streaming real-time thought processes directly to your console.

Installation 📦

Install globally via npm to make the mramma command available anywhere on your system:

npm install -g mramma-cli

Quick Start 🚀

To start the interactive coding agent, navigate to your project directory and run:

mramma

On your first run, you'll be greeted by the Setup Wizard, which will help you configure your preferred AI provider and API key.

Providers & Models 🤖

mramma supports an extensive list of top-tier AI providers. During setup (or anytime via /provider), you can select:

• Anthropic: Claude 3.7 Sonnet, Claude 3.5 Haiku, etc.
• Google Gemini: Gemini 2.5 Flash, Gemini 3.1 Flash Lite Preview, Gemini 2.5 Pro.
• GitHub Models: GPT-4o, Llama 3.3 70B, Mistral Large (Free tier available with a standard GitHub PAT).
• OpenRouter: Gemma 2 9B It, Qwen 3 Coder, and thousands of community models.
• Nvidia API: DeepSeek v3.2, Mistral Large 3, GLM 4.7.

Commands & Shortcuts ⌨️

Inside the mramma REPL interface, you can use several slash commands to control the agent:

• /config or /settings - Open the configuration wizard to change models, providers, or toggle bash Auto-Accept.
• /model - Quickly switch to a different model for the current provider.
• /mode - Toggle between Agent (full power), Chat (read-only), Fast (quick answers), or Plan modes.
• /tier - Adjust the default prompt execution tier (quick, build, architect, smart).
• /todos - View the active task list the agent is currently working on.
• /resume - Load your last session's context.
• /clear - Wipe the current session context.
• /compact - Compress the current context budget to save tokens.
• /cost - Display a detailed token usage and cost table for the current session.
• /diff - View Git changes made in the working directory during your session.
• /snapshot - Create a quick git stash of your current progress.
• /rollback - Revert to the last mramma snapshot.
• /update - Automatically update mramma-cli to the latest npm release.
• /exit or quit - Leave the agent.

You can also run raw shell commands directly by prefixing them with ! (e.g., !npm run test or !git status).

Prompt Tiers 🎚️

Mramma uses dynamic prompt tiers to match the effort required for your task. You can prefix your prompt with the following to enforce a specific tier for that turn:

• !q <prompt> (Quick): Fast, low-iteration task execution. Best for small code edits.
• !b <prompt> (Build): Standard robust task execution.
• !a <prompt> (Architect): Deep, multi-step planning and sub-agent delegation.
• !s <prompt> (Smart): Automatically evaluates your prompt to decide the best tier.

Environment Variables ⚙️

You can optionally configure mramma via environment variables (such as through a .env file or exported variables):

• MRAMMA_PROVIDER
• MRAMMA_MODEL
• ANTHROPIC_API_KEY
• GEMINI_API_KEY
• GITHUB_TOKEN
• OPENROUTER_API_KEY
• NVIDIA_API_KEY

License 📄

MIT License.