Introducing Craft Agents — The Open Source Agent Interface

We built Craft Agents because we wanted a better, more opinionated way of working with the most powerful agents in the world.

Not another terminal. Not another code editor with AI bolted on. A real interface — document-centric, beautiful, and fluid — designed from the ground up for how agents actually work.

Today, we're releasing it as open source.

Why we built this

Working with AI agents today often means living in the terminal, writing config files, and wrestling with OAuth flows. Every new integration is a project. Every context switch breaks your flow. The tools are powerful, but the experience isn't.

We think agents deserve better interfaces. Interfaces that feel native to how agents work — with multitasking, long-running sessions, and seamless connections to your tools.

So we built one.

Just tell it what you need

Want to connect Linear? Say "add Linear as a source."

Craft Agents finds the API, reads the documentation, configures credentials, and sets everything up — automatically. No config files. No setup wizards. No copying tokens between browser tabs.

The same works for Slack, Gmail, GitHub, Stripe, or any service with a public API or MCP server. Paste an OpenAPI spec, a URL, or even a screenshot of documentation — the agent figures it out.

Document-centric, not code-centric

Most agent tools are built for developers writing code. Craft Agents is built for anyone working with information.

Sessions feel like documents, not chat logs. You can flag important conversations, assign workflow statuses (Todo → In Progress → Needs Review → Done), and manage multiple parallel tasks in an inbox-style interface.

Drag and drop images, PDFs, and Office documents directly into conversations. They're automatically converted and ready for the agent to work with.

Built with Craft Agents

Here's something we're proud of: we're building Craft Agents with Craft Agents.

No code editors. Every feature, every fix, every improvement — just prompts. If we want to customize something, we ask. The agent handles the rest.

This isn't a demo. It's how we actually work. And it proves that agent-native workflows aren't just possible — they're better for many tasks.

You control the autonomy

Three permission modes let you dial in exactly how much freedom the agent has:

Mode	Behavior
Explore	Read-only. Safe research and investigation.
Ask to Edit	Approves each action before execution.
Execute	Full autonomy for trusted workflows.

Start safe, build trust, then let it run. Toggle between modes with a keyboard shortcut — Shift+Tab cycles through them instantly.

Workspaces and Skills

Workspaces keep your contexts separate. Personal projects, client work, different environments — each with its own sources, credentials, and conversation history.

Skills are reusable instructions you define once and invoke whenever you need them. Create a skill for code reviews, weekly reports, or content workflows. Mention it in conversation and the agent follows your playbook.

Native diagram rendering

We also built beautiful-mermaid — a fast, beautiful Mermaid rendering engine with no external dependencies. Craft Agents renders diagrams natively, so you can visualize architectures, flows, and schemas without leaving the conversation.