Bolt

You will like Bolt for the first hour: a prompt becomes a running app without installing anything, because the whole runtime lives in your browser tab. That is the trait that decides it, and it is a prototyping trait: the second day, when the app needs auth and a database, you are stitching services on by hand and the tab starts to feel small.

Pick it for small front-end experiments, demos you will throw away, and showing a stakeholder what you mean before lunch. Pick Lovable when the app needs a database and auth, which come with it there rather than as an afternoon of wiring.

The meter punishes size. The pricing page states that most token usage goes to syncing the project's file system to the AI, and that larger projects consume more tokens per interaction, so the bill climbs exactly as the app becomes real. The cheapest prompt is the one on an empty project; the dearest is the fix you need on the app you have shipped.

Watch the per-message cost after week two, and keep dead files out of the project. What it does right: WebContainers execute in your browser tab, so a bad command cannot reach your disk.

The substrate deserves credit: executing a Node runtime inside the browser tab isolates the agent's commands from the host without a server or a container. Above it, context is gathered by syncing the project file system to the model, which the vendor's own pricing page identifies as the dominant token cost; this is retrieval by volume rather than by relevance, and its cost grows with the repository rather than the task.

No benchmark is claimed, and no model is named for the agent. The consequence is that quality cannot be attributed to scaffold or model. The observation: the engineering below the agent is better documented than the agent.

The tell I dislike is the repository: the MIT core at stackblitz/bolt.new was published and then left alone, open source as a marketing plan, retired once it had done its job. The moat is WebContainers, a real technical asset; everything above it is contestable by Lovable, v0 and Replit, who compete on the same prompt-to-app promise with deeper backends.

The likely acquirer is a hosting vendor that wants the runtime and will fold the builder into a dashboard. Position: use it, export to GitHub often, and do not let the project live only in the tab.

The demo is a running app in a browser tab, which the design team will love. Procurement: Teams is $30 per member, so $1,800 a month for sixty; SSO, audit logs and data governance are Enterprise, which means a sales call before the security questionnaire. The meter grows with repository size, since most tokens go to syncing the project to the model, so a real codebase costs more than the demo did.

JavaScript only and no headless mode, so nothing runs in CI. Not yet for engineering; a small Teams plan for design is a different memo.

The stackblitz/bolt.new repo is MIT and frozen since December 2024; the thing at bolt.new is proprietary, Claude only, and takes no API key of mine, which is the pet peeve made flesh: open source with the good part behind a login, and the open part no longer moving. Nothing runs offline, since the model is theirs.

What I can bend: it is an MCP client, and two-way GitHub sync gets the code out so I can finish it in an editor I own. A fork of the repo is an artifact of 2024, useful for reading how WebContainers are driven and nothing else. Grudging respect for the runtime.