Skip to content

Cud

Simplicity is the Ultimate Foundation

Cud (named after the bolus of food that llamas chew) is a lightweight, local-first multi-agent framework. Built from the ground up to eliminate artificial complexity, Cud brings autonomous, tool-calling AI agents directly to your workstation or private infrastructure without third-party API dependencies or hidden prompt abstractions.

⚡ Key Features

Simplicity First

Zero convoluted setups or bloated abstractions. Built to be intuitive, inspectable, and easy to hack on.

Local & Private

Powered by Ollama. Your prompts, source code, and long-term agent memories stay 100% on your machine.

Multi-Agent Architecture

Deploy fleets of specialized autonomous agents, each with isolated personas, state, and sandboxed workspaces.

Transparent Markdown Prompts

Agent behavior and system prompts are fully transparent, written in plain AGENT.md files.

Tool-Rich Ecosystem

Native support for persistent bash sessions, filesystem operations, SKILL.md packages, and Model Context Protocol (MCP).

Daemon-Ready

Run agents as background Linux system services via systemd with active Discord gateways and scheduled tasks.

🖼️ Interface & Workflow Gallery

Explore Cud in action across its runtime environments and user interfaces:

=== "System Flow" Cud Interaction Flow

=== "Desktop GUI" Cud Desktop GUI Dashboard

=== "Agent Settings" Agent Configuration Panel

=== "Instructions Editor" AGENT.md Prompt Editor

=== "Memory Management" Long-term Memory View

=== "Skills & Capabilities" Agent Skills Manager

=== "Periodic Tasks" Scheduled Tasks Monitor

=== "MCP Integrations" Model Context Protocol Manager

=== "Subagents Delegation" Custom Subagents Configuration

📚 Documentation Index

Navigate through the comprehensive technical documentation for Cud:

Architecture & Core Design

High-level system design, Ollama LLM integration, DeepAgents execution loop, and state machine graph.

Agent & Workspace

Deep dive into ~/.cud/agents/<name>/ directory structure, AGENT.md, MEMORY.md, and history.db.

User Interfaces (CLI, TUI & GUI)

Command line interface, Rich/prompt_toolkit TUI, and PySide6 Qt Desktop application overview.

Built-in Tools & Skills

Shell execution, virtual filesystem backends, context compaction, and custom Markdown SKILL.md definitions.

Model Context Protocol (MCP)

Connect agents to external MCP tools via stdio or HTTP/SSE transports with secret environment injection.

Custom Subagents

Hierarchical agent delegation patterns, context isolation, and subagent-specific MCP servers.

Discord Gateway & Daemons

Connect agents to Discord servers with systemd background service integration and thread session management.

Scheduled Tasks

Autonomous cron-scheduled agent jobs configured using Markdown frontmatter and prompt templates.

Installation & Setup

Quick start installation script, manual pipx / uv tool setups, system requirements, and Ollama configuration.