How can we help?

Type a plain-English description of what you want — “an internal tool that tracks support tickets,” for example — and the agent spins up a private workspace, writes a plan, and starts building. You don’t need to scaffold anything yourself. See the docs for a full walkthrough.

No. You can start as a guest and the session is tied to this browser while your projects are stored securely — save the one-time recovery code from the guest banner to reopen them on another device or browser. When you create an account you can carry that guest project forward, so nothing is lost. Signing in also unlocks reopening projects across devices and multiple concurrent workspaces.

Before the agent changes anything, it can investigate your codebase and lay out a plan in plain English — a “what you’ll get” list and a wireframe up front, with every technical detail (files, steps, success criteria) one click away behind a “Technical details” toggle. It keeps the AI out in the open instead of silently rewriting your project. Plan is one of four permission modes — alongside Ask before edits, Auto-accept, and Full autonomy — and you can switch between them mid-conversation as the work changes shape.

The Free plan needs no credit card and includes a one-time $3 model wallet for trying real work. See pricing for BYOK, Plus, and configurable Max wallets.

Choose BYOK at $8/month when you want model charges on your own provider accounts. Plus is $20/month with $12 in build credit and a separate $2 retry/correction reserve for an immediate follow-up when work comes back broken. Max lets you choose $100–$200/month with the exact wallet shown before Stripe checkout.

Yes, on BYOK, Plus, and Max. BYOK requires Anthropic for every session because it powers internal planning and compaction; manually selected models also need their provider key. On Plus and Max, a configured key overrides included credits for that provider. Keys are encrypted at rest, never displayed back in full, and never exposed to the agent.

Guest projects are stored securely, while access to the guest session is tied to this browser or your recovery code. When you create an account, you can bring that work forward so you can reopen it anywhere and keep iterating. Nothing is thrown away when you upgrade from guest to a full account.

Enterprise customers can request a Data Processing Addendum, invoicing, and a security review. We also work with early-stage startups, nonprofits, and educators — tell us about your team through the Enterprise page.

Each project runs in its own isolated virtual machine. When the agent starts your dev server, the preview panel shows the running app and reloads as files change — so you see the real thing, not a mockup, while you build.

That’s almost always a dev server bound to 127.0.0.1, which isn’t reachable across the workspace boundary. Bind it to 0.0.0.0 instead — most frameworks accept a --host 0.0.0.0 flag or a host setting. Asking the agent to “make the preview work” will also nudge it to rebind for you.

Workspaces pause when idle and resume exactly where you left off, so you don’t lose state between sessions. Your code and chat history stay with your account, and you can delete a project at any time to remove its sandbox and data.

No. Each project gets a dedicated VM with its own filesystem and network boundary — not a shared container or browser tab. A runaway script or risky command in one project can’t reach another.

You can pick from a curated list across four providers — Anthropic (Claude), Z.ai (GLM), OpenAI (GPT), and Google (Gemini) — in the composer’s model picker, or set an account-wide default in Settings. The default is Auto, which sizes up each step and routes it to whichever configured model fits best — a fast model for a quick edit, a stronger reasoner for a hard problem — so you don’t have to think about it.

Thinking effort is a slider from low up through max — the exact rungs offered depend on the model you’ve picked, since not every provider supports the very top end. Higher effort is better for tricky, multi-step changes; lower is faster for small edits. It’s a per-turn control in the model picker with a default in Settings, and it applies across all four providers.

Yes. Capped Claude search works with Gate 15 credit or your own key. GPT, GLM, and Gemini 3.x use their native search when you provide that provider’s key. This keeps tools without a provider-side usage cap from silently spending the shared wallet. Searches appear as a clearly labeled activity step, and the system only advertises search when the selected model and credential mode support it.

It depends on the plan. Free uses Gate 15 keys while its one-time build credit remains. BYOK always needs your Anthropic key for internal work and the matching provider key for a manually selected model. On Plus and Max, your key overrides included credit for that provider; a blank provider uses Gate 15 credit until the build balance is exhausted.

Link your repository from the project, and Gate 15 keeps your workspace and GitHub in sync. Pushes are explicit actions you trigger — code never leaves your workspace until you say so.

You can ship your app with a one-click deploy when it’s ready. Like GitHub pushes, deploying is something you choose to do — the agent doesn’t publish anything on its own. You stay in control of when your app goes live.

Add them as encrypted secrets in your project. The agent can use a secret through controlled tooling so your app works, but the raw value is never printed back into chat or exposed to the model. Keys are encrypted at rest.

Yes — plan mode lets you see what the agent intends to do, and anything destructive or hard to undo is flagged for your approval rather than pushed through quietly. Combined with checkpoints, you always have a way back.

Yes. Projects save checkpoints as you work, so you can rewind to an earlier state if a change goes wrong — without losing the rest of your progress. Roll back to the last good point and try a different approach.

Give it a more specific instruction, or switch to plan mode so you can steer the approach before it acts. Bumping thinking effort up toward high or max helps on gnarly, multi-step problems. You can also try a different model from the picker — each provider has its own strengths.

Check that the dev server is bound to 0.0.0.0 rather than 127.0.0.1 — a localhost-only bind is the most common cause of a 502 or an empty preview, because it isn’t reachable from outside the VM. The activity log shows the exact command the server started with.

Reach out through the contact page with your project and what you were trying to do. Sharing the plan, the failing step, and any error from the activity log helps us help you faster.