Chris Tate

Introducing json-render AI-generated UI. Deterministic output. 1. Define your component catalog 2. AI steams JSON 3. Render interactive UI Let users prompt dashboards, widgets and apps - safely constrained to components and actions you define

ICYMI: portless get Print the URL for a service, useful for wiring services together BACKEND_URL=$(portless get backend) Applies worktree prefix detection by default. Use --no-worktree to skip it.

@coskar4422 Yes, the preview in the UI lets me manually test dev server(s) running on each sandbox

@ctatedev Never locally in your own browser? Basically do you ever do manual testing or all done by agents/automated tests?

~100% of my dev is done in sandboxes in the cloud Highly recommend it: - Unlimited parallel agent sessions - My local machine stays safe - Can work from anywhere - Can close laptop - Lap stays cool Interesting idea to visualize with Kanban

Quality code takes time to generate, at least 5m for each agent turn, and can easily take 30m+ per task Doing work in parallel means not waiting around. Flow is fire off several agents, review code for the first that finishes, (usually) send a follow-up, then round robin that all day

is it too late for me to again ask what you all are doing that requires *unlimited agents* running in parallel churning out code?

@coskar4422 Web UI with preview, code diffs and agent-browser

@ctatedev How do you do QA?

@Ibra_aka_ajax The example uses stdio but it should work with a remote server

@ctatedev amazing. stdio support only right?

Introducing Generative UI for MCP Apps One server. Infinite interfaces. Instead of building views, define a component catalog. The AI assembles the right UI based on your API, CLI or MCP server tools. Works in Claude, ChatGPT, VS Code, Cursor and more

@elie2222 @nbaschez I've been thinking a lot about this lately

@nbaschez @ctatedev Ya third party services has been the real challenge for us :(

@tobiastornros @ryancarson Yes, I do this for monorepos primarily

@ctatedev @ryancarson Does it work on larger repos? For example 10+ apps and 100+ packages.

@arthurliebhardt @ryancarson They are short-lived. One sandbox per task/PR. Usually 5-30m each. Separate, no host They start from base snapshot with tools / system packages (kinda like an image), then it just clones, install dependencies and runs the CLI agent

@ctatedev @ryancarson Oh ok, then I have a completely different image of what sandboxes are. I thought it's something like a "short-lived" Docker container that executes something. Is this more like a snapshot Unix system you are running here in Sandbox?

@arthurliebhardt @ryancarson Cursor CLI (and other coding agents) via UI

@ctatedev @ryancarson sound sick but somehow difficult to visualize for me when you say cursor on the sandbox you mean the cursor cli or does a Sandbox also have a UI?

@kirylch Yeah an evolution of it x.com/ctatedev/statu…

@ctatedev Do you use something like this template github.com/vercel-labs/co…

@willyzimmm @ryancarson You're right - this is more about what needs to be orchestrated There's def some good out of the box solutions This is a roll you own solution I use bc I really like organizing tasks in a way that makes it easier to work across a lot of repos, issues, PRs, releases at once

@ctatedev @ryancarson How does this help facilitate agent orchestration vs just doing it w symphony in codex/claude code? @ctatedev

@kentcdodds Yes, exactly. You can make an MCP server that exposes the renderer and your catalog as a tool Then the AI generates a spec, the host renders it, and users interact with the live UI

Looks awesome! Is this built with the MCP Apps spec? In ChatGPT at least it seems that your MCP server needs to list out available templates (resources) at connection time (so you can't generate those on demand). Do you have a generic shell template and then the tool accepts the json which is passed to that app?

@lmon_25 Ngl it's an awesome feeling

@ctatedev Need to start doing this. I find myself having to close my laptop mid prompt and it’s really annoying.

@Tyler_Palmer9 Great name 😆

@ctatedev Going to put this out soon for an agent kanban board. Working in diffs and accept/reject with the terminal. Something is dislike vs working in Cursor. It’s all local for now for security but eventually will look at a cloud runner for users.

My preference is spin up a fresh db per sandbox. I'd rather have *everything* be local External integrations can be tricker. For auth, I'll usually have email/pass instead of 3rd party, even if it's a local-only option For APIs, you can sometimes get away with mock but AI makes it really easy to code up a drop-in replacement so it's at least production-like

@ctatedev I wanna do this - the things I haven’t solved yet are anything tired to the URL (webhooks, oauth redirects, etc) and databases (spin up a fresh db per sandbox? Or point to one central dev db and somehow deal with conflicting migration state)

@jeffrschneider Amazing 🙌 I'm about to share something soon wrt one of the ways I use agent-browser in this setup

@ctatedev oh, I love the approach - I implemented it, that's how I know the license issue (and backed off). btw, fan of agent-browser, trying to embed it as well.

@PA80339646 Just 2 (external) keys: GitHub + whichever CLI agent I wanna use

@ctatedev How many API keys at play?

@jeffrschneider True, I'm not sure how that works tbh. This setup can work with any provider/model tho

@ctatedev no, licensing issue. It "works" fine on E2B, etc.

@entrecurious @ryancarson Both are mentioned here vercel.com/docs/vercel-sa…

@ctatedev @ryancarson What are you using for the Firewall + Credential Injection?

Yes - Vercel Sandbox This is the flow I've settled on for now. A lot automated by triggers And I've built a few web apps to manage + orchestrate it (might release something one day) - One agent session + PR per sandbox - Base snapshot: pre-built with Claude, Codex, Cursor, agent-browser, portless, etc. already installed - Automatic lifecycle management with snapshots and hibernation - Firewall + Credential injection: creds never enter the sandbox - Agent execution via CLI, detached - Logs and events consumed by ingress server (with poll fallback) - Clone of main branch and PR/feature branch for e2e, regression testing