Craft Agents is an open-source desktop application built in TypeScript that provides a document-centric interface for working with AI agents, specifically designed around the Claude Agent SDK and Pi SDK. The project is classified across multiple domains including autonomous agents, AI orchestration, multi-agent systems, agent communication, task automation, workflow management, and distributed intelligence. It was created by the team at craft.do to enable more intuitive multitasking and seamless integration with external APIs and services through a beautiful, fluid UI rather than traditional code-editor or CLI workflows.
The core philosophy behind Craft Agents centers on agent-native software principles, where users describe what they want and the system figures out how to accomplish it. The application handles API and service connections with minimal configuration friction. Users can tell the agent to add sources like Linear, Gmail, or Slack, and the system automatically discovers public APIs and Model Context Protocol servers, reads their documentation, sets up credentials, and configures everything without requiring config files or setup wizards. The platform supports multiple connection types including MCP servers, REST APIs from providers like Google and Microsoft, local filesystems, and custom APIs via OpenAPI specifications or endpoint URLs.
The desktop application includes comprehensive session management with an inbox and archive system organized by workflow status, flagging for important sessions, and AI-generated or manual session naming. Sessions persist with full conversation history saved to disk. The interface provides a multi-session inbox with status workflow capabilities, streaming responses with tool visualization similar to Claude Code, and real-time updates. Users can connect multiple LLM providers simultaneously, including Anthropic, Google AI Studio, ChatGPT Plus, GitHub Copilot, and OpenAI API keys, with per-workspace default settings. The platform includes 32 plus Craft document tools through MCP integration, allowing access to blocks, collections, search, and task functionality.
Permission modes operate at three levels: Explore for read-only access, Ask to Edit for approval prompts, and Auto for automatic approval, with customizable rules. The system supports background tasks with progress tracking, dynamic status workflows with customizable states, cascading theme systems at app and workspace levels, and multi-file diff viewing in VS Code style. Skills function as specialized agent instructions stored per-workspace, while file attachments support drag-and-drop for images, PDFs, and Office documents with automatic conversion. Automations enable event-driven agent session creation triggered by label changes, schedules, or tool use.
The project supports both desktop and headless server deployment modes, with the server capable of running on remote machines while the desktop app functions as a thin client. A CLI client enables terminal-based interaction for scripting and CI/CD integration. The codebase is licensed under Apache 2.0, allowing free remixing and customization, with the team itself building Craft Agents using Craft Agents rather than traditional code editors.
