Run multiple AI agent tasks in parallel using a visual inbox that shows each task's status from To Do through Done.
Connect the agent to Slack, Linear, or other tools by describing it in plain language, letting the agent set up the integration itself.
Run long tasks on a remote headless server and check in from any device using the desktop as a thin client.
Set up Automations that trigger agent sessions automatically when a label changes or on a recurring schedule.
Requires an API key for at least one AI provider such as Anthropic, Google AI Studio, OpenAI, or GitHub Copilot.
Craft Agents is a desktop application for working with AI agents, built by the team behind the Craft document editor. Instead of a terminal interface, it offers a visual inbox-style layout where you create and manage multiple agent sessions the way you might manage tasks, with status labels like To Do, In Progress, Needs Review, and Done. The core idea is that you describe what you want in plain language, and the agent figures out how to do it. Connecting to external services follows this same pattern: you tell the agent to add Slack or Linear as a source, and it locates the relevant API or MCP server, reads the documentation, sets up credentials, and configures everything itself. You can also paste an existing MCP config JSON or an OpenAPI spec and the agent handles the rest. Local tools, like a Python script or an npx command, are also supported as local subprocess connections. The app supports multiple AI providers in the same workspace: Anthropic (via API key or a Claude Max subscription), Google AI Studio, ChatGPT Plus, and GitHub Copilot. Each workspace can set its own default. Responses stream in real time with visible tool activity, similar to what the Claude Code CLI shows but inside a graphical window. File attachments including images, PDFs, and Office documents can be dragged in directly. Skills are saved instruction sets stored per workspace, and Automations let you trigger agent sessions in response to events like label changes or schedules. A permission mode system controls how freely the agent can take actions: read-only exploration, ask before editing, or auto-approve everything. A keyboard shortcut cycles through these modes mid-conversation. The app can also run in a headless server mode on a remote machine, with the desktop acting as a thin client, keeping long-running sessions alive and accessible from multiple computers. Installation is a one-line curl command on macOS and Linux, a PowerShell command on Windows, or a manual build using Bun. The project is open source under the Apache 2.0 license.
← craft-ai-agents on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.