Describe it
Write the project in plain English. Gate 15 names it, sharpens the first prompt, opens the workspace, and starts new projects with a plan you can review.
Gate 15 is an AI engineering workbench. You describe the outcome; the agent edits real files in an isolated sandbox, runs commands, starts previews, and reports back in chat. This guide walks the loop end to end — from a first prompt to a deployed app.
However you begin, projects reopen with their sandbox, chat sessions, previews, skills, secrets, and checkpoints intact.
Write the project in plain English. Gate 15 names it, sharpens the first prompt, opens the workspace, and starts new projects with a plan you can review.
Upload a .zip, clone from GitHub, or paste a repo URL with a token for a one-off import. Imported projects keep their file tree and chat history.
Projects reopen with their sandbox files, chat sessions, previews, skills, secrets, and checkpoints still attached.
The new-project card on the dashboard has three tabs. Pick the one that matches how much source material you already have.
Best for a new app, website, tool, automation, or prototype.
Write the project in plain English. Gate 15 names it in about 200ms and the workspace opens with your brief forwarded to the agent verbatim — new projects start with a plan first. The fastest path when you can describe the shape.
Best for existing source code on your machine.
Imports archives up to 250 MB compressed. The importer skips .git/, node_modules/, and build output (.next/, dist/, build/) so the sandbox starts clean.
Best for repos you want to edit, run, and optionally push back.
Connect GitHub to pick from your repos, or paste a URL and personal access token for private one-off imports. Guest accounts need to sign in before using GitHub.
Alongside your project list, the sidebar has account-wide tools you’ll use across every project.
Start a new project from a working example instead of a blank page.
Connect a Supabase project in a couple of clicks — Gate 15 provisions the database and stores the keys — then browse tables, preview data, and run SQL without leaving the dashboard.
Generate a full design-token set — colors, type, components — from a brief, an existing project, a live URL, a Figma file, or an upload, with a live preview styled in your own tokens.
Save reusable skills at the account level — curated starter packs or your own — and mark any to auto-apply to every new project. Separate from a project's own Skills doc under Configure.
Attach PDFs, spreadsheets, and Word docs as reference material the agent can read across your projects.
Invite teammates to a project as owner, admin, editor, or viewer, leave comments, and share an organization workspace.
The workspace is split like a lightweight IDE: conversation on the left, project files beside it, and the editor, preview, and logs on the right.
Give instructions, review plans, attach references, switch chat sessions, pick a model, and stop an in-flight turn.
Browse the sandbox tree, open files, edit directly, and reference files in chat with @-mentions.
Inspect code, preview Markdown, and view running apps in the preview tabs created by Run or the agent.
Watch command output from installs, builds, tests, and dev servers without leaving the workspace.
Run the app, deploy, create a GitHub repo, edit Skills, manage Secrets, rewind checkpoints, and toggle panels.
Check connection state, active project, sync freshness, branch label, and encoding at a glance.
Chat is where intent becomes code. The best turns tell Gate 15 what outcome matters, what evidence to inspect, and how to verify the result.
PlanAsk Gate 15 to inspect the project with read-only tools and propose editable steps before it changes files — you watch it investigate in real time. New projects start this way by default.Ask before editsPause for your approval before each edit, command, or risky operation.Auto-accept editsEdits and routine commands run right away; the agent still pauses for anything dangerous or expensive.Full autonomyRun everything without asking, with no safety prompts. Switch between any of the four modes at any time, even mid-turn, from the composer.FilesAttach images, PDFs, CSVs, design references, or other files the agent should use as evidence.@ fileReference exact project files so the agent reads the right code before editing./ commandsRun built-in or project slash commands from the composer.ModelStay on Auto — which routes each step to whichever configured model fits best — or pin Claude, GLM, GPT, or Gemini for this turn and dial thinking effort from low to max.StopCancel the current turn. Work already written to disk is kept, so you can redirect from there.Use the session dropdown to create a separate thread for a new investigation. Sessions keep their own conversation history while sharing the same project files and configuration.
Clear removes the conversation history for that session. It does not delete sandbox files, so the code remains available.
For multi-part builds, Gate 15 can spawn sub-agents that work concurrently on different pieces. The Activity tab shows live token/cost stats, sub-agent progress, and your task list as a turn runs.
Ask for an image and Gate 15 generates or edits one directly, then uses the result immediately in your project.
Click Run to start or restart the project dev server. The first run installs dependencies when needed, then opens a preview tab.
You and the agent share the same sandbox. Direct edits, agent edits, file opens, and command output all stay in sync.
Use the editor for tiny text tweaks, quick copy changes, or inspecting a file while the agent works. Save is automatic through the workspace sync.
Use @-mentions when you want Gate 15 to follow an existing pattern, patch a specific file, or explain why a piece of code behaves a certain way.
Some settings belong to a project, and others apply account-wide. Use project tools for instructions and secrets; use Settings for defaults and appearance.
Project instructions prepended to the agent's system prompt. Use them for conventions, design rules, preferred libraries, testing habits, and project-specific constraints.
Encrypted API keys and tokens stay outside the coding sandbox and model context. Trusted connectors and deployment backends resolve them server-side.
Set Auto or choose a specific provider/model in Settings — Anthropic, Z.ai, OpenAI, or Google. You can also override the model and thinking effort per turn from the chat composer.
Account-wide instructions and default skills for new projects. Use this for preferences you want every fresh project to inherit.
Switch dark/light theme and comfortable/compact density in Settings. Docs, dashboard, and workspace share the same tokens.
Signed-in accounts can connect GitHub, create private repos from workspace projects, and deploy apps to Vercel from the topbar.
Use Create GitHub repo in the workspace topbar to create a fresh private repo and push the initial project state.
Use Deploy to publish through Vercel. Set any environment variables in the deploy modal, then watch status and history in the same flow.
Download the whole project as a zip, drop your live app into any page with an embed snippet, or share a preview link that expires and can be revoked — no sign-in required for the person viewing it.
The workspace is built for iteration. You can stop an agent turn, start a cleaner chat, or restore a checkpoint when a direction goes sideways.
Interrupt the active turn and redirect from the current files.
Browse checkpoints and restore an earlier project state when a direction goes sideways.
Use the status bar to see whether files have been saved recently before you act.
Strong prompts are specific enough to act on and small enough to verify. Treat each turn like a useful ticket for a careful engineer — and know how the agent already works so you can lean on it.
Build a billing dashboard for a small SaaS founder so they can spot failed payments quickly. Use Next.js and the existing design tokens. Start with the overview screen, seed realistic mock data, run the app, and verify desktop and mobile layouts.
Make me a dashboard.
GoalSay who it is for and what success looks like.ContextName the stack, files, data shape, screenshots, constraints, and examples.ScopeAsk for one coherent slice, then iterate after you see it.ProofAsk for tests, preview checks, screenshots, or command output.Build [what] for [who] so they can [goal]. Use [stack or constraints]. Start with [first screen or workflow]. Use [attached files or @file references] as source material. Run [tests or preview] and tell me what you verified.
These are the working principles the Gate 15 agent already follows. Knowing them tells you what to ask for and what you can trust it to do on its own.
After meaningful frontend work, Gate 15 starts a preview and drives it — clicking through real flows, filling forms, submitting, navigating — then checks desktop and mobile, console errors, and accessibility before reporting done. You can ask for this explicitly: “run the signup flow and confirm it lands on the dashboard.” You'll watch each step live in a Preview (Agent) tab.
Add API keys and tokens in the Secrets pane. They remain encrypted outside the coding sandbox and resolve only inside trusted connectors or the deployment backend. Don't paste live credentials into a normal message.
You don't have a terminal in the sandbox, and the agent knows that — it won't tell you to run npm run dev, installs, builds, or deploys yourself. If a command is needed, it runs with the agent's tools and the output streams into Logs. When a web app is ready, the agent shares the real public preview URL, never a localhost link.
The preview reaches your dev server across a network boundary, so a server bound to 127.0.0.1 shows up as a blank 502. The agent binds to 0.0.0.0 for you. If you bring your own start script, make sure it passes the framework's host flag (next dev -H 0.0.0.0, vite --host 0.0.0.0, flask run --host=0.0.0.0, and so on).
Training data lags reality by months, especially for AI model names, library versions, and prices. When your task depends on the current lineup — a model picker, a “compare the latest” page, a freshly released API — say so. The agent can web-search to confirm before it writes code instead of trusting a stale memory.
Most problems get easier when you hand the agent the exact symptom and the evidence. These are the gotchas that come up most.
Almost always a dev server bound to localhost instead of 0.0.0.0 — the preview proxy can't reach it across the sandbox boundary. The agent binds correctly by default; if you supplied a custom command, tell Gate 15 the preview is 502 and to rebind the server to 0.0.0.0 and restart. An empty server log with a 502 is the tell.
Open Logs and ask Gate 15 to read the failing command output, fix the root cause, and rerun the app. Common culprits are a wrong port, a project in a subdirectory, or a syntax error surfaced in the log. Paste the exact error if you see one.
This is a dependency-install race — two installs running in the same folder at once. Let the agent handle installs; it reconciles package.json and the lockfile at the sandbox root or the subdirectory being started. If you see missing-module errors after a manual install, ask Gate 15 to reinstall cleanly and restart the server.
Long sessions hit the model's context limit and get compacted (summarized) automatically, which can blur older detail. For a clean break, start a new chat session — it shares the same files, skills, secrets, and sandbox but begins with a fresh thread. Re-state the key facts the agent needs at the top.
Gate 15 reconnects on its own; the status bar shows connection state and how recently files synced. If a turn looks stuck after a reconnect, refresh the workspace — your sandbox files and chat history are persisted, so nothing is lost. Then ask the agent to confirm the current state with git status or by re-reading the file.
Stop the turn, point at the specific file or UI area, and ask for a focused correction. Use Rewind to restore an earlier checkpoint when a direction goes sideways.
Connect GitHub from the dashboard or Settings, or paste a token with repo access. Guest accounts must sign in before GitHub features appear.
Attach screenshots, name the target screen size, and ask Gate 15 to verify with the preview. Concrete, specific visual feedback lands far better than asking it to make it nicer.
Start with $3 in trial usage. BYOK from $8/month. Included model credits from $20/month.