penecho

Introduction: Think with AI beyond the chat box. A shared canvas for handwriting, equations, diagrams, and spatial reasoning.

More: Author ReportBugs

Tags:

A spatial workspace
for thinking with AI.

Draw, explore, and build with the built-in Agent or your own MCP-compatible assistant.

Website · Download · Quick start · MCP guide · Discord

PenEcho full demo PenEcho professional diagrams demo

PenEcho plugins demo PenEcho interactive canvas demo

Keep talking in Codex, Claude, Kimi, or other AI agents. Let PenEcho give the work a place to live.

Through MCP, your AI can turn explanations into diagrams and ideas into interactive previews. Keep references, reasoning, and work side by side — then mark up the canvas and bring feedback into the next round.

Keep the conversation	See the work take shape	Bring feedback back
Work with the AI agent you already use.	PenEcho's MCP server brings diagrams, documents, and interactive previews onto the Canvas.	Try the result, annotate it, and let your agent read your feedback for the next revision.

An architecture discussion, annotated by hand on the Canvas.

Draw professional diagrams that are easy to explore and interact with.

Architecture

Kubernetes production cluster	Monolith-to-microservices migration
Sequences

Event-driven notifications	How MCP requests reach the Canvas
Workflows

Release preparation: parallel tasks	Multi-region rollout and rollback

_{Click an image to view the full-size diagram.}

See it before it’s finished. See the work take shape as you talk with AI. Try it, give feedback, and move your project forward together.

Connect your agent with MCP →

What you can do

Work visually. Combine handwriting, equations, text, images, diagrams, and interactive HTML Widgets on a spacious canvas.
Create with AI. Use the built-in Agent to research, work with files, explain ideas, and create editable visual results.
Bring your own agent. Connect Codex, Claude Code, or another MCP-compatible client to read and edit an explicitly enabled Canvas.
Keep and share your work. Organize Canvases into projects, save Cloud revisions, sync favorites, and publish through Echoes.

New in 1.3.3

Diagram	What you can draw
Architecture diagrams	Model services, dependencies, and nested system boundaries with automatic layout and routed connections.
Sequence diagrams	Show participants and message order, including replies, self-calls, and conditional, loop, or parallel fragments.
Workflows	Map steps, decisions, labeled branches, loops, and parallel paths with fork/join points.

Describe your requirements to PenEcho Agent or an MCP-connected agent. Inspect details on the Canvas, refine the result through feedback, and export SVG or PNG.

Full changelog

How it works

PenEcho architecture: a browser connects to PenEcho Cloud or your local PC. Cloud includes hosted models and connects to your linked device. Your PC runs PenEcho CLI or App with your LLM API or agents. External AI agents can use Cloud MCP or Local MCP; both connections are optional.

Open PenEcho in a browser through PenEcho Cloud or your local PC running the CLI or desktop app. Cloud provides hosted models and can connect to your linked device; your PC can use your own model API or agents. External AI agents such as Codex and Claude can connect through Cloud MCP or Local MCP. Both MCP connections are optional.

See the architecture notes for implementation details.

Quick start

Desktop: download the Windows or macOS app from GitHub Releases.

npm: requires Node.js 22.19 or newer.

npm install -g penecho
penecho

Open http://localhost:3888. Add your own model API or an authenticated Codex, Claude Code, or Kimi CLI in Settings → Connections. Connections are saved in ~/.penecho/connections.json; general settings remain in ~/.penecho/config.env. For PenEcho-hosted models, sign in and select an available model in Settings.

At startup, set a six-digit access code or explicitly enable open access on your trusted network. Startup also prints LAN addresses for other devices.

Run from source

git clone https://github.com/penecho/penecho.git
cd penecho
npm install
npm start

Connect your agent with MCP

For Local MCP:

Start PenEcho and enable the current Canvas in Settings → MCP service.
Use Settings to configure a supported local client or copy its generated launch configuration. For a global npm installation, clients that accept mcpServers JSON can use:
```
{
  "mcpServers": {
    "penecho": { "command": "penecho", "args": ["mcp"] }
  }
}
```
Ask your agent: “Show the architecture we discussed on my PenEcho Canvas.”

The agent can capture relevant content, edit objects, create visual results, patch document source files, and receive your feedback. Only enabled, connected Canvases are discoverable. With Local MCP, the MCP client runs on the PenEcho host; support for LAN and linked-device browsers does not expose the local MCP endpoint publicly. Cloud MCP is a separate authenticated HTTPS connection for your enabled PenEcho Cloud canvases.

Desktop installations should use the generated configuration, which includes the correct bundled runtime. See MCP setup and the optional agent workflow skill.

PenEcho Cloud and AI connections

PenEcho Cloud adds private versioned projects, synced favorites, public sharing through Echoes, and remote access to a linked computer.

Connection	How it works
PenEcho models	Sign in, select an available hosted model, and use account credits. Settings shows current rates and balance.
Your model API	Configure an OpenAI- or Anthropic-compatible endpoint, model, and API key. Usage is handled by your provider.
Your CLI	Use a locally installed and authenticated Codex, Claude Code, or Kimi CLI. Availability and usage depend on that provider's plan.

Hosted models on your computer require a Cloud sign-in, without device pairing or a separate Credits API key. Cloud MCP can access enabled Cloud canvases directly. Accessing a Canvas hosted on your computer through Cloud requires the linked device to be online and the necessary relay support.

Your own API and CLI connections do not spend PenEcho credits. A Cloud account is optional for local use with your own connection. AI features require access to the selected provider; running PenEcho locally does not make a remote model available offline.

Recommended model configurations

These recommendations balance answer quality against the latency of PenEcho's real canvas workload, based on current hands-on testing; actual response time varies with the provider, canvas complexity, and reasoning behavior.

Model	Effort	Notes	Recommended use
Claude Opus 4.8 / 5.0 (`claude-opus-4-8` / `claude-opus-5-0`)	`medium`	Strong quality with a better latency balance	Everyday canvas work
Claude Opus 4.8 / 5.0 (`claude-opus-4-8` / `claude-opus-5-0`)	`high`	Higher reasoning quality, longer and more variable waits	Complex handwriting, mathematics, diagrams, or layout
Fable 5 (`claude-fable-5` or `fable`)	`medium`	Often around half the response time of `gpt-5.6-sol` at `xhigh`	Fast, high-quality general use
Kimi K3 (`kimi-k3`)	`medium`	Very good quality; `medium` keeps the balance practical	Recommended Kimi default
`gpt-5.6-terra`	`low` to `high`	Surprisingly strong and responsive	Flexible quality and latency targets
`gpt-5.6-luna`	`xhigh`	Very good canvas results with strong speed	Quality-first, still responsive
`gpt-5.6-sol`	`high`	Good enough for most requests, more responsive than `xhigh`	Default when responsiveness matters
`gpt-5.6-sol`	`xhigh`	Very good but slower and more variable	Difficult canvas tasks
`deepseek-v4-flash-vision-exp`	`medium`	Good	Vision-capable work through the DeepSeek API
`glm-5.3-flash`	`medium`	Good	Fast work through the GLM Anthropic-compatible API

Community and license

Read CONTRIBUTING.md to contribute; run npm run check before opening a pull request. Report bugs in Issues, discuss ideas in Discussions, or join Discord.

Licensed under AGPL-3.0-only. Alternative commercial licensing is available. See the trademark policy and contributor agreement.

Acknowledgements

Thanks to Archify by tt-a1i. PenEcho’s professional diagram renderers use adapted SVG and geometry helpers from this MIT-licensed project. The MIT license and copyright notices are preserved; see NOTICE for third-party attribution.