Claudable: a Next.js builder that shells out to your local CLI agent

The gap Claudable is trying to fill

Prompt-to-app builders normally run the model on someone else's infrastructure. You get a browser tab, a monthly bill, and code you cannot run until you export it. Claudable inverts that. The README describes it as a Next.js-based web app builder that combines Claude Code's agent capabilities with Lovable's app building experience, and the operational consequence is that the model call happens on your machine through a CLI you already installed and logged into. The stated benefit is that there is no additional subscription cost for the builder itself. The pitch is aimed at developers who already hold a Claude Pro, Max, Team or Enterprise plan, a ChatGPT Plus or higher plan, a Cursor subscription, or who are willing to run Qwen Code for free. If you fall into one of those groups, the marginal cost of the builder is zero and the generated files land in a project directory you control. If you do not, Claudable has nothing to drive, because it ships no model of its own.

How the agent handoff actually works

The architecture visible in the README is a Next.js application that spawns a locally installed CLI agent as a subprocess and streams its output back into a preview pane. The README lists the supported agents as Claude Code, Codex CLI, Cursor CLI, Qwen Code and Z.AI GLM-4.6, and each has its own install command and login step, which means Claudable is not abstracting over a uniform API. It is driving five different command-line tools that happen to share a shape: a binary on PATH, an authenticated session, and a working directory. The generated output is Next.js, styled with Tailwind CSS and shadcn/ui according to the feature list, and the preview is described as hot-reload so changes appear as the agent writes them. Deployment goes to Vercel and the database layer to Supabase. That is the whole data flow: prompt in, CLI agent writes files, Next.js dev server reflects them, and two external services handle hosting and persistence when you are ready. There is no intermediate representation and no vendor-hosted build step.

Getting it running: three commands and a login

The prerequisites section is short and worth reading literally: Node.js 18+, Claude Code or Cursor CLI already logged in, and Git. The quick start is a clone, an install and a dev server. The README gives the commands as git clone https://github.com/opactorai/Claudable.git, then cd Claudable, then npm install, then npm run dev, with the app available at http://localhost:3000. It notes that ports are automatically detected and that if the default is in use the next available port is assigned, so do not hardcode 3000 into a script that expects a fixed URL. Before any of that you need an agent on PATH. For Claude Code the README gives npm install -g @anthropic-ai/claude-code followed by claude and then /login. For Codex CLI it gives npm install -g @openai/codex and codex, logging in with a ChatGPT account. For Cursor CLI the install is a curl piped to bash and then cursor-agent login. Qwen Code installs with npm install -g @qwen-code/qwen-code@latest and is verified with qwen --version. Z.AI GLM-4.6 points at an external quick start guide rather than an inline command.

The agent you pick changes the economics, not just the output

Because Claudable is a thin layer over five CLIs, the choice of agent is the choice of billing model, context window and licence. Claude Code is listed as included with Claude Pro, Max, Team or Enterprise plans, or an Anthropic API key, with a native 200k token context. Codex CLI is included with ChatGPT plans from $20 per month. Cursor CLI has a free tier and a Pro tier from $20 per month on a credit-based system. Qwen Code is described as completely free and open-source under Apache 2.0, with 256K native context and up to 1M with extrapolation, and model sizes from 0.5B to 480B. Z.AI GLM-4.6 runs from $3 per month for GLM Coding Lite to $30 per month for GLM Coding Max, with a first-month discount. The README marks Claude Code as recommended. That recommendation is doing real work: the 200k context and the described deep codebase awareness matter when the agent is editing an existing Next.js project rather than generating a greenfield file. A smaller-context agent will hit the point where it stops reading the whole tree sooner.

No sandbox is the design, and also the risk

The feature list advertises zero setup, no API key and no database headaches, and explicitly says there are no complex sandboxes. Read that as a description of the trust model rather than a convenience. A hosted builder runs generated code inside an isolated container it controls. Claudable runs the agent on your machine, in your filesystem, with whatever permissions that CLI was granted. The README does not describe a permission prompt layer, a container, or a review step before files are written. Codex CLI is described as having interactive, auto-edit and full-auto operating modes, which tells you the agent itself decides how much autonomy to take, not Claudable. If you run a full-auto mode in a directory that contains anything other than the generated project, the blast radius is your filesystem. The practical mitigation is to clone Claudable into a directory you are willing to lose, and to start the agent in its most conservative mode. That is a workflow decision the README leaves to you.

Where it stops being the right tool

Two limitations are visible from the material alone. The first is that everything depends on a local, authenticated CLI session, so Claudable cannot run in a browser-only environment, a locked-down corporate laptop that blocks global npm installs, or a CI runner without credentials. The second is upgrade friction. The troubleshooting section addresses database migration conflicts by telling users who upgraded from a previous version and hit errors to reset the Prisma database so it matches the latest schema. A reset is a destructive operation. If you have data in the local database, the documented remedy costs it. There is also a mismatch worth flagging: the repository is listed under anymorph-ai/Claudable while the README's own links, including the license and star badges, point at opactorai/Claudable. That is a redirect or a rename, and anyone writing automation against the GitHub URL should confirm which one resolves before pinning it. The single listed release, v1.0.0 from August 2025, is described as fully CLI-powered multi-agent, and the most recent push is April 2026, which suggests active work between releases rather than a frozen snapshot.

What you would use instead, and why the difference matters

The obvious comparison is Lovable, which the README names as the experience it is imitating. The difference is not cosmetic. Lovable hosts the generation, so you get a shared project URL, a browser-only workflow and a bill that scales with usage, but you do not get to choose the model or keep the agent session on your hardware. Claudable inverts all three: model choice is yours, the session is local, and the builder itself is free under MIT, while collaboration and the browser-only path are gone. A second comparison is running the CLI agent directly in a terminal and skipping the builder. That is strictly more flexible, and it is what an experienced Next.js developer will do most of the time. Claudable's added value is the preview loop and the Vercel and Supabase wiring, which is worth something when you are iterating on UI rather than on architecture. If neither the preview nor the deploy integration saves you time, the wrapper is overhead.

Licence, maintenance and what the material does not cover

Claudable is MIT-licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. That covers the builder. It does not cover what the builder produces or what it calls: Qwen Code is Apache 2.0, and the other agents are proprietary services governed by their own terms, so the licence of your generated application is a separate question from the licence of Claudable. The README's own note that there is no additional subscription cost and built just for YOU is a claim about the builder, not about the agents. On maintenance, the material supports only a narrow statement: one release, v1.0.0, and a last push in April 2026. There is no changelog, no upgrade guide beyond the Prisma reset advice, and no stated support policy. The Electron desktop builds for Mac, Windows and Linux are listed as a feature but the README gives no build or download instructions for them, so treat the desktop path as unverified until you find that documentation yourself.

Editorial conclusion

Adopt Claudable if you already pay for one of the supported CLI agents, work on a machine where you can install Node.js 18+ and Git, and want prompt-to-preview without a hosted builder account. Do not adopt it if you need a browser-only workflow, a team-visible audit trail, or a guarantee that a generated app is safe to run: the README does not describe any sandboxing, so the agent's own permission model is your only boundary. Before committing, verify three things on your own machine: that your chosen CLI is logged in and reachable from the shell Claudable spawns, that port 3000 is free or that you accept the auto-assigned fallback, and that the Prisma schema matches your Node version, because the troubleshooting section only offers a database reset when migrations conflict.

Official sources

1. anymorph-ai/Claudable on GitHub
2. Issues
3. License: MIT
4. README
5. Releases