Sketch by Hand, Let Your AI Agent Do the Rest
You've probably had this experience: you're working through a system design or a data flow, and the fastest way to think it through is to grab a pen and start sketching. Arrows, boxes, scribbled notes. Then you want to actually use that thinking—turn it into a real diagram, a document, or a working prototype—and you're stuck redrawing everything from scratch in some other tool. PenEcho is a project that tries to close that gap by putting a canvas at the center and letting AI agents read what's on it.
What It Does
PenEcho is a desktop app (and a web version at penecho.ai) built around an endless canvas where you sketch by hand. You can write text, draw equations, drop in images, and jot down notes. The app watches what you're doing and—when you pause—a built-in Agent can respond automatically. It takes your handwritten input and turns it into something structured: architecture diagrams, sequence diagrams, workflow diagrams, documents, or interactive widgets that actually run on the canvas.
The interesting architectural choice here is that PenEcho doesn't lock you into its own AI. It ships with a built-in Agent, but it's also MCP-ready, which means you can connect external tools like Codex, Claude Code, or any MCP client. So the canvas becomes a kind of shared workspace between you and whatever AI you prefer to use. Diagrams come with automatic layout and can be exported as SVG or PNG. Widgets—calculators, quizzes, prototypes—run directly on the canvas and can be favorited or shared.
It's released under AGPL-3.0-only, currently at version 1.3.5, and available on npm as well as through direct downloads for macOS and Windows.
Why It's Cool
-
The sketch-first workflow is the whole point. Most diagramming tools make you think in their UI first—drag this shape, connect that line. PenEcho flips it: you think however you naturally think (messy handwriting, half-finished arrows), and the AI handles the translation into structured output. That's a meaningfully different interaction model.
-
Auto-answer on pause is a subtle but clever touch. You don't have to explicitly ask for help every time. You sketch, you pause, and the agent responds. It treats the canvas less like a document you submit and more like a conversation you're having.
-
The MCP support is the real unlock. Being able to plug in Claude Code or Codex or any MCP client means you're not betting on PenEcho's built-in agent being the best one forever. The canvas stays useful as the AI landscape shifts underneath it. That's a design decision that respects how fast this space moves.
-
Widgets that actually run. A lot of tools stop at generating a diagram image. PenEcho generates things that execute—calculators, quizzes, prototypes. That's a bigger leap than it sounds, because it turns the canvas from a documentation surface into a working surface.
-
Export options that don't trap you. SVG and PNG export means your diagrams can leave the app and go into a README, a slide deck, or a design doc without a fight.
How to Try It
The fastest path is the web version—just open penecho.ai and start sketching. If you want the desktop app:
- Head to the releases page.
- Download the build for your platform—macOS or Windows.
- Install and launch. No account setup mentioned in the README, so you should be able to dive straight in.
If you'd rather work from the command line, PenEcho is published on npm as penecho:
npm install penecho
To connect your own AI agent instead of (or alongside) the built-in one, the project has a dedicated setup guide at docs/mcp-setup.md. The README advertises MCP readiness as a first-class feature, so if you're already running Claude Code or Codex locally, that's probably where you want to start.
The repository lives at github.com/penecho/penecho, and there's a Discord if you want to ask questions or follow development.
Final Thoughts
PenEcho is aimed at people who think spatially and hate the friction of translating a whiteboard sketch into a polished artifact. If your workflow already involves sketching out architecture or logic before you build, and you're comfortable wiring up an MCP client, this could slot in nicely. The AGPL-3.0 license is worth noting if you're thinking about commercial use—it's a copyleft license, so check the terms if that matters to you. It's early days at v1.3.5, and the project leans heavily on the assumption that you want AI in your thinking loop. If that's you, it's worth a download.
Follow @githubprojects for more developer tools and open source projects.