PenEcho

Some ideas are spatial before they are verbal, an architecture, a workflow, a handwritten equation, and forcing them through a chat window loses that. PenEcho, built by Eric Kong and first released mid-2026, is a canvas app built around that observation: draw, annotate and arrange handwriting, diagrams and interactive widgets, while an AI agent (its own built-in one, or Codex, Claude Code or Kimi CLI connected over MCP) reads and edits the same surface. The pitch is explicitly a spatial extension of an existing AI conversation, not a replacement for it.

A local server (desktop app or npm install -g penecho) runs the canvas at localhost:3888, and MCP connects it two ways: your agent can capture what is on the canvas and turn explanations into architecture, sequence or workflow diagrams with automatic layout, and you can annotate the result by hand for the agent to read back on the next turn. PenEcho brings no model of its own; you connect your own OpenAI- or Anthropic-compatible API key, an authenticated Codex/Claude/Kimi CLI, or PenEcho Cloud's hosted models on credits, and Cloud MCP extends the same loop to canvases you have not enabled for local access.

It fits work that is genuinely spatial: explaining a system design and getting a routed architecture diagram back, sketching a migration plan and iterating on it visually, or working through math and diagrams that a chat transcript renders badly. Because feedback loops through the canvas rather than prose, it is aimed at people who think by drawing and marking up rather than by describing, and the two-way MCP connection (agent edits, you annotate, agent reads the annotation) is the feature that makes that loop actually work rather than being a one-way export.

AGPL-3.0 is genuinely OSI-open, but it ships alongside a separate commercial license and a mandatory Contributor License Agreement, a standard open-core structure, and an honest one, but it means contributions can be relicensed and the company can sell around AGPL's network-copyleft obligation, so check COMMERCIAL-LICENSE.md before assuming unrestricted SaaS use. Despite 2,392 stars and 314 forks, the project is effectively maintained by one person (43 of 44 commits), and because it brings no model of its own, output quality depends entirely on whichever provider you connect; the README's own model-recommendation table is already a snapshot that will age quickly as providers ship new versions.

PenEcho is worth trying if visual, spatial collaboration with an AI agent, diagrams, handwriting, iterative annotated feedback, is a real gap in your current chat-only workflow, and native desktop builds plus a real release cadence suggest it is more than a demo. The open-core license structure and solo-maintainer reality are the two things to factor in before depending on it for anything beyond personal use: read the commercial license terms if this heads toward a company SaaS deployment, and treat the model-quality bar as whatever you bring to it, not something PenEcho itself guarantees.