Upload screenshots, inspect a site, describe a UI, import code, open Figma or GitHub, or use Stitch directly. Every path lands in the same local project workspace.
Installer .exe recommended. Sign in to GitHub Copilot from the app — no command-line tool needed — or bring one API key or your own endpoint.
Generated code is model output: review it before running it outside the sandboxed preview. The published Windows binaries are not code-signed, so Windows SmartScreen will warn on first run — verify the SHA-256 checksum instead.
Your screenshots, the generated code and your API keys stay on the machine you installed it on.
Upload, URL, Text, Import, Figma, GitHub and Stitch each expose the controls and safety boundaries that input needs. Multiple screenshots can become separate pages, responsive views, UI states or references; website inspection creates full-page evidence plus DESIGN.md; Import accepts fixed built Storybook metadata; repositories can open locally or run a selected-model first refinement.
The Code tab is a multi-file project tree with an editor, file tabs, Entry, Preview, Read-only and Generated badges, and a whitespace-only Format action. Current code is what you edit; a read-only Export project view shows the exact files a download will contain. Select an element in the preview and describe the change.
On wide windows, drag the divider between Chat and the preview, or between the file tree and the editor. Both are keyboard-operable separators, both remember your width across restarts, and both are clamped so neither side becomes unusable. A pane width is a view preference — resizing never creates or changes a version.
Projects, versions, prompts and conversations live in local SQLite under %LOCALAPPDATA%\shot2code\ — outside the installation folder. History expands every version's models, options, prompts, activity and ancestry; Full history searches every saved project before opening it. Upgrades keep the database, and deleting a project removes it from this device.
Export a self-contained single HTML file, or a Vite project folder for the stack you chose. Multi-file source projects are exported as they are rather than wrapped in an invented scaffold.
You still need a model provider. shot2code does not ship a model. Use Copilot, a Gemini/Anthropic/OpenAI key, or Copilot SDK BYOK. Its local Ollama preset needs no paid inference API, but Ollama, hardware and a vision-plus-tool-calling model with its own licence remain yours to install and operate.
Choose the path that matches the evidence you already have. The dedicated input-tab guide covers credentials, limits, privacy and output in detail.
DESIGN.md directly, or explicitly convert that evidence through selected models.
DESIGN.md and bounded full-page responsive evidence carry actual dimensions and explicit blank/truncation state before generation.
iframe.html.Current packaged-app captures of Review, permissioned integrations, provider-neutral tools, Chat inventory and local History. Select any screenshot to open its original pixels.
An agent loop with real tools, running from your machine against the provider you configured.
Choose Upload, URL, Text, Import, Figma, GitHub or Stitch. Pick the output stack where one applies, add an optional first instruction, and select models only when the path needs generation or conversion.
The model calls tools — create a file, edit a file, extract assets, render a screenshot preview — and shot2code executes them locally and feeds the results back.
One option is generated per selected model, in parallel, so you pick the closest one. Then say what to change, or select an element in the preview and describe the edit. Chat keeps the whole branch — your prompts, the attachments, the model identity and the assistant's own answers.
Every version is saved locally and reachable from History or Recent projects. Export a single HTML file or a Vite project folder when you are happy.
Generation streams over a local WebSocket from a bundled FastAPI backend; the desktop shell starts it on a free port. The only outbound traffic is to the model provider you chose, the update check, and the integrations you switch on and use — an MCP server you trusted, a Figma, Stitch or GitHub import, public-site inspection, opt-in web or free-image search, an image provider, feedback you explicitly submit, or anything you explicitly share.
What the app does today — and which capabilities need an extra key.
http://localhost:11434/v1; local inference needs no paid API, but Ollama, hardware and a separately licensed vision-plus-tool-calling model remain yours.gh or Copilot CLI session signed in.read_web_page, CC0/Public Domain Mark Openverse photos, and fixed-origin sanitized Iconify SVGs. Built-in whole-page web_fetch remains blocked.DESIGN.md, SKILL.md and a palette image.Ctrl+/ opens the Help centre, which carries the full shortcut reference. Ctrl+1–4 switch Preview, Code, Chat and History; Ctrl+3 again collapses the chat panel on wide windows, and Ctrl+Alt+C brings it back. Project actions use Ctrl+Alt combinations — N new, I import, U upload, S settings, E export. The same commands sit in the native File, View and Help menus, greyed out until a project makes them possible. Pane dividers take arrow keys, Shift for larger steps, Home/End and Enter to reset. In the desktop app, Ctrl+= and Ctrl+- zoom in fixed steps between 50% and 300%, and Ctrl+0 resets.
CodePen is offered only when the selected stack can honestly run in a browser-only Pen, and it always asks before code leaves the device. For build-dependent projects, download the project folder instead.
When export cannot find a valid root build command, it keeps every source file and adds a Safe fallback note rather than inventing a scaffold that looks plausible and breaks later.
Only the core routes have to be ready before the window opens — generation, evaluation and project tooling load on first use, and the Chromium and Copilot checks run in the background. The shell demands a real health response rather than any reply, gives up after 90 seconds instead of hanging, and logs how long readiness took.
Drag the divider between Chat and the preview, or between the file tree and the editor. Arrow keys move it, Shift moves it further, Home and End jump to the limits, Enter resets and Escape cancels a drag. Widths persist, clamp to the window, and never touch a project's versions.
Ctrl+/ opens a Help centre — Get started, Guides, Support and the full shortcut reference — linking the published guides rather than a private copy. In the desktop app it also opens the backend log folder, which is the first thing to attach to a bug report.
Each one has an explicit single-HTML and project-folder export strategy, so what you download actually builds.
HTML and CDN projects export as a Vite project with a deterministic static copy, so inline modules and import maps are not silently reordered. React and Preact use Vite's framework build path when the generated source is safely transformable, and a documented Vite HTML fallback when it is not.
One provider is enough — and whichever you bring, you choose which of its models actually run. Keys are stored on your device and sent only to the provider they belong to.
The easiest option, because there is no API key to manage and nothing to install: Sign in with GitHub in Settings runs the app's own GitHub OAuth device flow — a one-time code, your browser, a public client id and no client secret, and a token encrypted by the operating system key store. Running gh auth login (or copilot) still works, and disconnecting inside shot2code leaves those CLI sessions signed in. Needs an active Copilot subscription, and reaches Claude, GPT, Gemini and Grok models through it. The model list is discovered live from the signed-in account, so it is your plan's real entitlement.
Paste an API key into Settings. It is kept on this device and sent with the generation request to that provider only. Each of these providers has a maintained, validated model list, so picking models for them contacts nobody. Gemini additionally powers video input and asset extraction.
Key setupAn optional Copilot SDK BYOK connection reaches an OpenAI-compatible, Azure OpenAI or Anthropic endpoint of your own. No Copilot subscription is required. It uses its own API key or bearer token; direct OpenAI/Anthropic keys are never borrowed for it, only an OpenAI-compatible endpoint on localhost may run without one, and there is no Gemini equivalent. Configure local Ollama fills http://localhost:11434/v1. shot2code does not install Ollama or a model: local inference needs no paid API, but uses your hardware and a separately licensed model with image input and tool calling. Native providers are never rerouted.
Choose Replicate, Cloudflare Workers AI, or an OpenAI-compatible image endpoint in Settings. Credentials stay independent from the model providers. Replicate remains the default-compatible choice and the only background-removal provider; editing appears only when the provider really supports it.
What needs which keyChoosing models is part of Settings, not a Copilot-only extra. Settings → Models — and the compact picker beside the composer — group every model you can currently use under GitHub Copilot, OpenAI, Anthropic, Google Gemini and, when you configure one, Copilot SDK (BYOK). Tick as many as you like, across providers, and each run produces one option per selected model, up to 4 for a first generation and 2 for an update or a video; tick nothing and the choice stays automatic. Models that cannot read the input you gave — video in particular — are filtered out rather than offered and failed, deprecated models stay hidden unless you ask for them or already picked one, and a saved model that can no longer run is reported and skipped instead of silently deleted. A provider only appears once the app can actually see a credential for it.
A BYOK option is its own selection, not a redirect. Each model your endpoint can serve is listed separately as sdk-byok/<provider>/<model>, so you can tick a model and its BYOK counterpart and get both results from one generation — at the same reasoning effort, side by side. The four native groups are untouched by it: each still needs its own key or sign-in, and a native pick always runs on its native provider. History records the identity each option actually ran as, so a retry repeats the same runtime. An incomplete or switched-off connection is reported as a notice and never blocks a direct generation.
MCP servers extend the SDK runtimes only. Up to eight servers over stdio, HTTP or SSE, added by hand or found through the in-app browser for the official MCP Registry — where installing an entry creates a disabled, untrusted draft for you to review, as does the featured Google Stitch server. Figma MCP endpoints are filtered because Figma restricts them to catalog-listed clients. A server must be enabled and explicitly trusted, stays read-only until write tools are allowed, a local server is spawned as an argument vector rather than through a shell, and a remote one must use https:// unless it is on localhost. Its tools reach Copilot/BYOK options only; a native OpenAI, Anthropic or Gemini option never sees them.
Skills remain SDK-only; canonical research does not. Search snippets, separately consented bounded page reading, Openverse photos and Iconify icons reach every runtime. The Chat Tools inventory reports their state, generated-image billing/readiness, MCP write scope and enabled Skills without turning anything on. Iconify's public API is currently keyless but has no fixed quota/SLA; built-in web_fetch stays blocked because its whole result cannot be inspected before the model receives it.
The latest stable release closes the gaps around local models, bounded research, design-system inputs, complete history, runtime Review and fail-closed recovery.
The existing BYOK card can configure http://localhost:11434/v1. Local inference needs no paid API, but Ollama, hardware and a separately licensed vision-plus-tool-calling model remain yours. A separate read_web_page permission reads public query-free pages inside byte/text/call budgets; runtime-owned web_fetch stays blocked.
Full history searches every local project and expands models, options, prompts, responses, activity, statuses and retry/branch ancestry. The GitHub tab now makes local open versus selected-model first refinement explicit, includes a design-system choice, and preserves the detected repository stack.
Built Storybook import reads only fixed JSON metadata from local files/folder/ZIP/public HTTPS and never executes stories, bundles or iframes. Figma frames can be previewed before the model call and reused. URL inspection scrolls bounded lazy content and reports real dimensions plus blank/truncation state, capped at 40,000px and 36 million pixels.
The Chat inventory reports consent, generated-image billing/readiness, MCP write scope and enabled Skills without enabling anything. Openverse remains the no-key CC0/Public Domain Mark photo path because general web-image results grant no reuse right. Iconify SVGs are fixed-origin, SPDX-filtered, sanitized, localized and provenance-tagged; its currently-keyless public access has no fixed quota or SLA.
Review combines source findings with bounded per-frame runtime evidence, categories, filtered select-all, isolated frame failures and explicit healthy/stale/partial coverage. Schema-v2 reports are not WCAG certification. Agent screenshot preview now returns nearly-blank state plus bounded console/page diagnostics.
Non-retryable image-provider failures trip a per-generation circuit breaker instead of repeating paid calls. At the application level, a truly blank renderer is reloaded once and then replaced with a static recovery screen while local project data remains intact.
Every version, its notes and all its files are on the releases page. The full history lives in the changelog.
Windows 10/11 x64. Each build carries its own Python backend and a headless Chromium, so there is nothing else to install. latest is the current published build; every version this hub has published — with the files it actually carries — is on the releases page.
Per-user NSIS install with shortcuts — and the only format that updates itself from GitHub Releases.
Download .exePer-machine deployment. Self-update is disabled by design; upgrades stay with the administrator.
Download .msiUnzip anywhere and run shot2code.exe — nothing to install, and easy to remove.
SHA-256 hashes for every file in this release. The binaries are unsigned, so this is how you check what you downloaded.
SHA256SUMSCompare the hash of the file you downloaded against the published list:
Get-FileHash .\shot2code-<version>-x64.exe -Algorithm SHA256Only the per-user .exe updates itself: Settings shows the running version, the update state and Restart & install once a download is ready. Before anything is replaced, the installed backend is stopped and that shutdown is verified — by the app, and again by the installer itself, so an update that starts from an older build is protected too. The installer now stops the installed backend itself before replacing anything, matching processes by path under resources\backend, never by name, and if it cannot confirm they are gone the install stops rather than overwriting a running app. MSI upgrades stay with the administrator and portable builds do not update.
Only what you ask for, and every route is yours to control.
Screenshots, generated code and Full history stay on the device. A generation goes only to the selected provider. Figma/Stitch/private-repository credentials remain capture-only. Local Ollama uses your own machine. Validating BYOK/MCP, opening Chat Tools and running source/runtime Review contact nothing; page/photo/icon research leaves the device only after its separate opt-in and a model call.
There is no analytics SDK. The remaining background call is the per-user update check. Everything else is an action you take: registry/design/Storybook import, AI review, web search, bounded page reading, Openverse/Iconify lookup, provider-billed image generation or confirmed CodePen sharing.
Imported projects/GitHub archives are parsed, never executed; Built Storybook reads only fixed JSON and never stories, bundles or iframes. Public pages, Openverse and Iconify pass public-address/type/size/sanitization bounds and remain untrusted evidence. Previews use an opaque origin plus restrictive CSP. MCP needs enablement and trust, writes need another switch, and Skill scripts cannot run.
Everything needed to install it, use it, work out what went wrong, and see what ships inside it.