cdbx: a browser-based IDE that aims to be both a toy and a tool

Browser-based IDEs generally fall into two camps: either a stripped-down “backlit notepad” or a heavy-duty cloud workspace that requires just as much configuration as a local machine. cdbx tries to straddle the middle ground—and on paper, it looks more promising than most of its competitors.

On one hand, there’s the Monaco editor (the same engine as in VS Code), a real terminal with output streaming, and instant previews without npm install. On the other—the “describe it in words, watch the app build” marketing pitch, a “publish” button, and zero barrier to entry. These are two different products packaged into a single interface, and that’s where the main risk lies.

MCP—Not Just a Buzzword, but a Working Combination

The most useful feature of cdbx isn’t “vibe coding,” but the MCP connectors. Planning Agent can query Notion, GitHub, Linear, Slack, and Google Drive on the fly, gathering context for a PRD or architectural document. This isn’t just a chat with an uploaded file: the AI makes requests to connected services on demand, without preloading anything. For teams whose backlog and documentation already live in Notion and Linear, this offers real time savings on copying context.

There are six plan types—General, Web App, Architecture, Deep Research, Slides, and API Spec—each with its own prompt and preview. After planning, a single click opens a new experiment with the context already loaded. The “think → build → publish” cycle feels seamless.

Code Execution API — a separate product

In addition to the IDE, cdbx offers an API for running code in an isolated sandbox container. 30 languages, one endpoint, responses in milliseconds. Network isolation, memory limits, seccomp profiles, a fresh container for every call. Free plan—100 runs per month; Pro—2,000.

Key point: The cdbx MCP server exposes run_code as a tool—Claude can call it directly. This means that the agent generates code and executes it immediately, receiving stdout, stderr, and an exit code. For “AI writes → AI checks” pipelines, this is more practical than deploying your own Docker runtime.

Where the Cracks Are

The Free plan provides 10 AI credits per month for Haiku and GPT-4o-mini. The Pro plan for $20 includes all models, 100–2,000 credits, and a monthly rollover. It sounds generous, but “credits” are a black box: it’s unclear how many credits a single Planning Agent cycle with five connector queries costs. If a single PRD burns through 20 credits, the Pro plan runs out before the middle of the month.

Bring-your-own-key for Anthropic and OpenAI partially solves the problem—you pay for tokens directly. But then why pay $20 for Pro if you have your own key? Just to avoid a badge on published apps and use custom domains? The cdbx team will need to explain more clearly where their added value ends and your infrastructure begins.

Verdict

cdbx won’t replace a local IDE for serious development—a browser is still just a browser. But as a prototyping environment with agents that actually read your context and run your own code, it’s more interesting than most “AI coding” wrappers. The question that will determine its fate: will they be able to retain both audiences—or will they have to choose one?

Source: Hacker News - Newest: ""AI" "LLM""