Craft Agents: a document-native Mac command center for AI agents
Craft Agents by Craft (Craft Docs Ltd) is an open-source Mac desktop tool that centralizes AI agents around a document-native workflow, intended to coordinate complex, multi-step tasks across a user's digital tools. The app connects local files and external services to each session, treats conversations as persistent documents, and keeps session data on the device. Knowledge workers, developers, and power users gain a single control surface for research and automation tasks.
SDK and protocol layers give agents direct access to external tools
The app is built on the Claude Agent SDK and adds support for the Model Context Protocol, which lets agents call third-party tools and servers. Zero-configuration discovery uses natural-language commands to link services, so the agent can reference repositories or chat threads without manual API wiring. Typical integrations mentioned include GitHub, Slack, and Linear, and the protocol support allows the agent to reach additional MCP-compatible services.
On-device encryption separates local data from external model queries
The tool stores conversation history and credentials locally and protects them using AES-256-GCM encryption, preserving a local-first privacy model. Model inference requires an active internet connection for LLM APIs. The app lists support for multiple providers, including Anthropic (Claude), OpenAI, Google AI Studio, and local models run through Ollama or LM Studio, so model selection happens at the connection layer rather than inside stored sessions.
Parallel sessions and controlled autonomy suit complex, multi-step work
Sessions act as manageable documents and the inbox-style workflow makes it practical to run parallel research or automation tasks without losing context. A built-in Chromium-based browser permits the agent to navigate pages, extract data, and capture screenshots during research. Permission-based autonomy exposes Ask to Edit and Full Auto modes; Explore mode is read-only while Execute mode lets the agent perform edits and send messages under the configured permissions.
Developer-oriented install and open-source code enable customization
Installation on Mac is available via a one-line terminal command; the app also supplies installation scripts for Linux and Windows. The project is published under the Apache 2.0 license, which supports community auditing and modification. The developer’s experimental division is associated with the project, and the open-source codebase makes it practical for technically minded teams to extend integrations or inspect how external connections are handled.
A capable tool for technically minded users who accept operational trade-offs
The app is a practical choice for developers and power users who need structured, agent-driven workflows and who can manage external model credentials and permissions. It rewards that investment with a programmable control surface and on-device session governance, though it expects deliberate oversight of autonomous actions and active model connections. Recommended.
