Open source CapCut alternative.
ChatGPT App
The ChatGPT app connects a Donkey Cut account and lets ChatGPT edit its cloud projects: import footage, inspect it, cut it, caption it, preview, undo, and export, with the preview and the export download playing inside the conversation. ChatGPT is the model; Donkey Cut checks ownership and runs the typed commands it sends.
The one rule: the ChatGPT adapter lives under site/src/clients/chatgpt.
Editing, rendering and playback stay in Cut; the adapter only translates.
ChatGPT tools and embedded card
↓
Scoped OAuth connection
↓
Cloud project, command and render services
↓
Command batch → the editor in the card runs it → versioned save
Captured document → worker renderer → signed media
Editing from ChatGPT
Each editor command is its own tool, annotated by its effect; the editor in the card runs it on the document the user sees as one undo step, saves, and reports. With no card open the tool asks for the project to be opened, and undo and redo step the card's own history, which the user's edits share. Editing, previews and exports spend no credits; commands that generate media spend the account's credits, and imports and exports use its storage allowance; the server instructions and tool descriptions say so.
Client boundaries
The standalone widget imports @donkeycut/artifact-player/ArtifactVideo, the
same player used by website previews and shared videos. The widget is
built by npm run chatgpt:build, which also runs before the site's development
and production builds. Rebuild it after editing widget code. ChatGPT caches the
widget HTML it reads over MCP, so the entry script and stylesheet keep stable
names that revalidate on every load; only the chunks they import are hashed.
The MCP adapter exposes projects, editor commands, guides, imports, history, previews, exports and job polling. Its schemas and widget data types share one contract. Signed media and download URLs travel in tool-result metadata, which reaches the widget. The model gets project metadata, command outcomes, render state, and an Open in Donkey Cut link. ChatGPT's sandbox drops null-valued keys from a tool result before the widget sees it, so the widget contract treats a missing key and null alike.
Opening a project plays the proxy the editor already rendered for it, so nobody waits on a render to watch. A render is for a revision with no current proxy, such as one edited from ChatGPT; it uses the existing headless renderer and a captured document, and repeated requests for the same saved revision reuse a queued, running, or completed preview. Only cloud projects reach ChatGPT.
The widget polls while visible and renews expiring playback URLs through the host bridge; it holds no account cookie or OAuth token.
Editable cards frame the editor or the projects home through a one-use,
one-minute link that creates a partitioned session cookie. Reconnect mints a
fresh link. Revoking the connection ends its sessions, and frameDomains permits the app's own domain.
Account linking
Donkey Cut supplies the account identity to ChatGPT through OpenID Connect.
The user signs in and consents to project access and identity sharing. The
openid and email scopes expose the account ID, email, and stored verification
status through /api/chatgpt/oauth/userinfo.
Discovery lives at /.well-known/openid-configuration and
/.well-known/oauth-authorization-server. Configure CHATGPT_OIDC_PRIVATE_KEY
with a PKCS#8 RSA private key of at least 2048 bits before deploying. Code
exchange signs an ID token bound to the client and requested nonce; the public
key is served at /api/chatgpt/oauth/jwks.
Consent binds the browser, account, callback, scopes, resource, and S256 challenge. Codes are single use, refresh tokens rotate, and replay revokes the connection. Tokens are hashed, MCP and UserInfo check expiry, revocation, resource and scope on every request, and rate limits and daily cleanup apply. A signed-out authorization is remembered, so signing in from any tab returns to consent. Revocation leaves already issued media links valid until their normal expiry.
Deployment and verification
The integration defaults to disabled in the chatgptApp setting. Deploy the
Prisma models in site/prisma/Chatgpt.prisma through the project's database
release process, deploy the site and cloud worker, then turn the setting
on.
Set the same CUT_RUNNER_SECRET on the hosted site and the
donkey-cut-worker Cloudflare Worker before deployment. The Worker passes it
to the render container.
Create a ChatGPT developer-mode app pointing at
https://donkeycut.com/api/chatgpt/mcp. Use the predefined public OAuth client
ID donkey-chatgpt, no client secret, and scopes
projects:read previews:render projects:write openid email.
Configure the exact redirect URI supplied by ChatGPT in chatgptApp.redirectUris.
Discovery advertises issuer identification, and every authorization redirect
includes iss. The embedded
resource declares https://chatgpt.donkeycut.com as its widget domain, from
which ChatGPT derives a sandbox origin. The widget HTML arrives through MCP,
its scripts and styles load from donkeycut.com, and video loads from
media.donkeycut.com; chatgpt.donkeycut.com needs no DNS record or Worker.
Public submission verifies control of donkeycut.com with a token from the
portal. npm run chatgpt:plugin -- --demo <url> builds the ZIP the portal
imports as a new draft: the listing, review cases and guides as skills.
Run the client tests with bun test src/clients/chatgpt. The browser fixture
bun run scripts/eval-chatgpt-widget.ts checks the widget against a mock
host. With the dev server and a local worker up,
bun run scripts/eval-chatgpt-workflow.ts drives the editing workflow through
the MCP tools in process.
Publishing changes
OpenAI keeps the tool metadata it scanned with each published version. A changed
tool description, schema, annotation, _meta, CSP, frame domain or server
instruction ships on deploy and replaces the published one once OpenAI's
automated checks pass; removed tools go at once. Widget code at the same
resource URI goes live on deploy, cached up to an hour. Renaming a tool, moving
the widget's resource URI, or editing the listing needs a new version submitted
for review, and a new server origin needs a new app. See OpenAI's
review guide.
Source entrypoints
site/src/clients/chatgpt/server/— OAuth, MCP tools, the command catalog, and the cloud-project adapter.site/src/clients/chatgpt/ui/— standalone MCP Apps widget and host bridge.site/packages/artifact-player/— shared website and widget playback.site/src/cut/server/cloud/commands.tsandsite/src/cut/lib/hostCommands.ts— the command job, and the editor in the card running it.site/src/cut/server/cloud/previewJobs.ts— revision reuse and document-backed rendering.
The protocol follows OpenAI's authentication guidance and MCP Apps UI guidance.