Toolcraft: a starter kit for building custom design apps with an AI agent

The problem Toolcraft targets: repeated manual work in a general design tool

The README frames the decision as a comparison, not a replacement. It says to use Toolcraft "only when a custom app would solve your task faster or better than Figma or another design tool," and to first account for what your existing software already does, including its AI features, plus the time needed to build the app.

The scenario it names is procedural graphics. Managing hundreds of small layers and reapplying effects or adjustments by hand gets slow when a brand's visual language is generated rather than drawn. A custom app lets you fix which controls and parameters exist, how the design responds to them, and which export formats you need, video included. The output is a repeatable workflow for producing consistent assets in batches, and something you can hand to a client so they generate new assets themselves.

That is a narrow audience. Toolcraft is for designers and small studios with a recurring asset pipeline, and for engineers who want a canvas-shaped starting point instead of wiring pan, zoom, layers and export from zero. It is not for one-off artwork.

What the create command actually copies into your folder

Toolcraft is not an application you log into. The README answers this directly: it is "a batch of files copied into a folder of your choice" when you run the command-line tool or ask an agent to do it. The repository confirms the shape with two top-level entries, starter/ and examples/, plus the licence and README. Everything the app becomes lives in your project afterwards.

The starter is a TanStack Router and React app in TypeScript, styled with Tailwind CSS and Base UI. On top of that sit the pieces that make it feel like a design tool: a canvas with pan and zoom, a component library covering buttons, dropdowns, sliders, font pickers, color palettes, curve editors and image uploaders, and optional layers, animation timelines and keyframes that an agent enables as the app needs them. Image and video export workflows are included.

The part that distinguishes it from a plain template is the AI harness. It supplies skills, instructions and checks that constrain an agent to Toolcraft's design system, and the repository lists unit, browser, acceptance and performance checks. In practice the harness is the product. A component library saves typing; a harness that tells an agent which components to use and how to verify the result is what keeps a 30-iteration build from drifting.

Installing Toolcraft and getting a first preview on port 3002

The README gives two routes. The agent route is a prompt you paste into Codex, Claude Code, Cursor or another coding agent, with the bracketed section replaced by your idea:

Install any missing prerequisites, including Node.js and npm. Set up Toolcraft
with `npx --yes @pixel-point/toolcraft create --yes`,
then build [describe your tool] and start a local dev server.

The agent installs Node.js and npm if they are missing, runs the create command non-interactively, and starts the dev server. Nothing else is typed.

The terminal route is one command, followed by prompts:

npx @pixel-point/toolcraft create

After the prompts finish, open the generated project in your agent and describe the tool. The README's own example is a request to "Build an app that applies an ASCII effect to an uploaded image and start a local preview." The dev server defaults to port 3002; if that port is busy on first launch, Toolcraft picks the next available one and remembers it for the project. Your agent usually reports a URL such as http://localhost:3002. If it does not, ask the agent for the local URL or to start the dev server.

To start from a finished example instead, the gallery at toolcraft.sh offers a Clone with CLI button on each example page that copies the setup command. The repository's examples/ directory lists the set, including ascii, dither-lab, draw-tool, mesh-gradient, particle-typography and liquid-metal. The README does not document a rollback or uninstall path, so treat the generated folder as the unit of cleanup.

Time and model spend are the real constraints

The README's own numbers are the most useful limitation. A first working result usually arrives 30 to 60 minutes after the initial prompt. Harder work, such as reverse engineering a reference, combining several WebGL effects, or building 3D tools, takes longer. Pixel Point says it often spends 1 to 3 days on a single app and goes through 30 to 50 iterations with AI, sometimes more, before the result matches the intent.

Budget follows from that. Asked whether a $20/month Codex, Cursor or Claude plan is enough, the README answers no, and says $20/month plans "aren't suitable for meaningful work," recommending a plan of at least $100/month. That is a recurring cost on top of the build time, and it does not shrink as the app matures.

The README also warns against expecting instant results: Toolcraft "isn't a magic tool that lets you create anything instantly." Where the documentation is thin is in failure modes. There is no published guidance on what happens when an agent cannot satisfy the acceptance checks, how to reset a build that has gone wrong, or how much of the harness you can safely edit. Those gaps matter because the harness is the part you are least likely to have written yourself.

Toolcraft versus a general design tool, and versus assembling the stack yourself

The honest alternative for many readers is the design tool they already pay for. Figma plus its AI features is the comparison the README itself raises, and it declines to argue Toolcraft wins by default. The difference in approach is where control sits. In Figma you manipulate layers and effects through a fixed interface; in a Toolcraft app you decide which parameters exist and how they map to output, which is what makes batch generation of brand assets tractable. If your work is a handful of distinct compositions rather than a parameterised system, Figma is the cheaper answer and the README says so.

The second alternative is assembling the stack directly: TanStack Router, React, Tailwind and Base UI, plus your own canvas, component library and agent instructions. That is the same foundation without the harness, the prebuilt controls, or the unit, browser, acceptance and performance checks. You trade Pixel Point's opinions for full control, and you write the agent scaffolding yourself. Toolcraft's value is concentrated in exactly that scaffolding, so if you plan to replace it, the starter kit is mostly a React template.

The README does not name any competing starter kits, so there is no like-for-like comparison to make here.

Licence, maintenance and what you inherit

Toolcraft is MIT licensed, and the README addresses commercial use twice: the software is free to use in commercial projects, modify and distribute to clients, and the starter repository's MIT licence permits commercial use, modification and distribution for paid client work. That is a permissive grant, and the practical consequence is that the apps you generate are yours to license as you choose. It is not legal advice, and if you redistribute the starter itself you should keep the licence file intact.

The repository is not archived, and its last push was on 2026-09-15, two days before this writing. That is the only maintenance signal available: no releases were retrieved, so there is no version history to judge cadence against, and nothing in the README describes an upgrade path for projects generated by an earlier version of the create command.

Upgrade cost is therefore the open question. Because Toolcraft copies files into your folder, later improvements to the starter, the component library or the AI harness do not arrive automatically. Keeping a generated app current means diffing your project against a fresh copy, and the README does not describe a supported way to do that. Plan on treating each app as a fork you own.

Editorial conclusion

Adopt Toolcraft if you are building a focused, canvas-based tool for a repeatable workflow and you already pay for a coding agent at the $100/month tier or above. Skip it if a general design tool already covers the task, or if you want a hosted product rather than a folder of files you maintain. Before committing, run the create command, confirm the port the dev server actually binds, and read the acceptance and performance checks in the starter to see what the AI harness will hold your agent to.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. pixel-point/toolcraft on GitHub
4. Project website
5. README