What I Learned from Building Open Canvas

I just released Open Canvas v1 beta — a browser-based workspace I built to solve a problem most people don't realize they have: their data management is the bottleneck holding back their coding agents, not the agents themselves.

Why I Built Open Canvas

Here is the origin story, stripped to the essential truth. I had been using coding agents daily — Claude Code, Codex, Gemini — and the agents were fast. Absurdly fast. I could describe a single-purpose application in plain language and have a working tool in minutes. Need a dashboard for a client engagement? Built. Need a chart on a dataset in a spreadsheet? Built. Need an HTML page to visualize email metrics? Built.

I got spoiled. Because the agents could build apps so fast, I found myself doing it for everything. If I knew I'd do a task more than twice, I'd just tell the agent to build a tool for it. A hundred clients? Build a hundred niche apps. It doesn't matter — the build time is negligible. The question was never “how long will this take to build?” It was always “do I have the data?”

The speed of generation made something obvious: building the tool was never the bottleneck. What slowed me down every single time was the same set of upstream problems:

• Scrambling to locate the data I needed and tell the agent where it lived
• Reformatting, reorganizing, or modernizing data so the agent could work with it
• Re-explaining the same context every time I started a new session
• Manually scoping the agent's working directory and permissions for each task

The agent could build a React app in three minutes, but I spent fifteen minutes getting it the right files, in the right format, with the right instructions. That ratio was wrong.

So I told Claude: build me a web app that solves this. And that was the first version of Open Canvas.

Data Is the Multiplier

Here is the core learning — the single most important thing I took away from this project:

The value you get from AI agents is directly proportional to how well you manage the data they operate on. Not how good the model is. Not how clever your prompts are. How accessible, organized, and persistent your data is.

Most people approach coding agents with ad hoc data management. They open a terminal, point the agent at whatever folder they happen to be in, manually paste context, and hope for the best. When the session ends, the context is gone. When they start a new session tomorrow, they reconstruct everything from scratch.

This is the equivalent of hiring a brilliant consultant and then making them re-read every document from the beginning of the engagement every single morning. No memory. No continuity. No compounding.

The fix is not a better model. The fix is a data strategy.

Two Principles

Open Canvas was built on two foundational ideas.

Centralize and Reuse Your Data

I organized all my projects within a single working repository. Each project lives in its own directory, but at the root level, there's a shared data layer. When I start a new project, I link to global shared data — client information, reference docs, templates — without duplicating it.

An agent working on Project A and an agent working on Project B can both access the same source-of-truth data. No redundancy. No drift. No “which version of the client brief is this?”

Dynamically Allocate Skills

In Claude Code, a skill is a set of reusable instructions that the agent can invoke automatically. Skills live in a .claude/skills/ directory and persist across sessions.

Open Canvas exploits the skill hierarchy deliberately. At the root level, it maintains a global skill that evolves with the workspace. Each project ships with default skills — including one called Dynamic Skills that simply says: “dynamically add skills as you see fit.”

The agent doesn't just execute tasks — it optimizes itself. As it works on a project, it identifies recurring patterns and creates new skills to handle them automatically. The agent gets better the more it works.

The Workflow Shift

Before Open Canvas, my workflow lived in VS Code. For every project, I'd open the folder, go to the terminal, run npm install, npm run dev, and switch to the browser. If it was Angular, Python, or a multi-stack app, the ceremony multiplied.

Open Canvas abstracts this behind a single pattern: every app includes a run.sh file. Whether it's a simple HTML file or a full React + Express stack, pressing “Run App” executes the same script.

The browser workspace gave me things VS Code couldn't:

• Unified view: App preview, terminals, and agent conversation in a single tab.
• Parallel agents: Multiple terminals running Claude, Codex, and Gemini on the same project simultaneously.
• Ambient monitoring: I can be in Gmail and just click the next tab to check what my agents have been doing.

With the right permission settings — like acceptEdits mode or the auto mode that uses a safety classifier — I can have ten Open Canvas tabs open, each with agents working autonomously. All I do is check in.

What I Actually Got Wrong

I thought the browser interface would be the main value. It wasn't. The interface is nice, but it's not what cut my time in half.

What cut my time in half was the cumulative elimination of data friction. The centralized data means I don't hunt for files. The skill hierarchy means the agent already knows how to handle each project. The run script means no setup ceremony. The permission config means I don't click “approve” forty times per session.

Every one of those friction points was small on its own. Together, they were consuming half my productive time. That is the data strategy tax I was paying without realizing it.

The Takeaway

The specific tool doesn't matter. What matters is whether you've thought about your data strategy at all. If you're using coding agents without one, you are reinventing the wheel every single session and leaving the majority of the value on the table.

Spend a day on it. Figure out where your data lives, how to centralize it, and how to configure your agents with persistent knowledge. Set up CLAUDE.md files. Create skills for repeatable workflows. Let the agents build their own expertise.

The agents are ready. The models are capable. The tooling exists. The question is whether your data is ready for them.

For the enterprise perspective on this topic, read the full insights article: Data Strategy: The Missing Layer in Coding Agent Workflows.

To learn more about Open Canvas, visit the Open Canvas product page.