Ask Claude, ChatGPT or Gemini for a product video and you get a script, a shot list and a prompt for some other tool — but not the file. This walkthrough shows how to connect an assistant to your media production over the Model Context Protocol so it can draft, revise and render inside editable JsonCut projects, with visual review before anything ships.
Why your chat assistant never finishes the job
Ask Claude, ChatGPT or Gemini for a product video and you get a script, a shot list, maybe a prompt for a separate image tool. The actual file — the thing you can post, embed or hand to a client — still has to be produced somewhere else, usually by you, one export at a time.
Short answer: connect your assistant to your media production over the Model Context Protocol (MCP) and it can work directly inside your projects: draft layouts, swap text and media, request review frames and trigger final renders — while every result stays an editable JsonCut project you can keep refining in Studio.
That last part matters. Plenty of pipelines can spit out a finished MP4. Far fewer keep the working file editable so a human can fix the one line of text that overflows — tomorrow, without redoing the whole job.
What MCP changes
MCP is an open-source standard for connecting AI applications to external systems — data sources, tools and workflows. The protocol's own documentation describes it as "a USB-C port for AI applications": one standardized plug that assistants such as Claude and ChatGPT, and developer tools such as Visual Studio Code and Cursor, already support.
For media work that means your assistant doesn't need a bespoke integration to touch your projects. JsonCut exposes its V2 authoring surface as a remote MCP server, so any MCP-capable client can:
| Capability | What your assistant can actually do |
|---|---|
| Editable projects | Create and edit layered image and video compositions instead of one-off files |
| Media | Upload or discover owned images, clips and audio and place them as layers |
| Templates | Search published templates, read their typed inputs and render them with your data |
| Visual review | Request review frames or a contact sheet before anything is final |
| Rendering | Render a pinned project version to PNG, JPEG, WebP, MP4, WebM, MOV or GIF |
The full setup is documented in the JsonCut MCP guide; the quickstart covers the basics of projects and renders if you're new to the platform.
Connecting: OAuth in five steps
For ChatGPT, Claude and other clients that support remote MCP with OAuth, JsonCut recommends the OAuth flow — no API key pasted into a chat window:
- Add a remote MCP server in your client using the endpoint
https://mcp.jsoncut.com/mcpand choose OAuth when asked for an authentication method. - Follow the sign-in link to JsonCut Studio and sign in with your usual account.
- Review the app name, return address and requested permissions, then allow access.
- Return to your client; it can now discover the JsonCut tools.
- Ask the client to list the JsonCut tools as a first check.
Connector availability and setting names depend on your client and plan. Each approved connection shows up as OAuth · [app name] in Studio's developer settings — delete that entry to revoke access. Access tokens last one hour and compatible clients refresh them automatically; after signing out of JsonCut or revoking a connection you simply reconnect.
For clients without OAuth, or for unattended automation, send a JsonCut API key as an X-API-Key header instead:
| OAuth (recommended) | API key | |
|---|---|---|
| Best for | Interactive use in Claude, ChatGPT, similar clients | Headless scripts, CI, clients without OAuth |
| Credential handling | Sign-in flow, nothing pasted | Secret header, scope it to what's needed |
| Lifetime | 1-hour tokens, auto-refresh | Until rotated or revoked |
| Revocation | Delete the entry in Studio developer settings | Delete or rotate the key |
One warning from the docs: an API key is not an OAuth bearer token, and you should never use a JsonCut API key as an OAuth client secret.
Verify you're on the right endpoint
After connecting, ask the assistant to call jsoncut_v2_get_authoring_guide. A correctly connected V2 server exposes only jsoncut_v2_* tools. If you see JSON job tools instead, the client is pointed at the frozen legacy endpoint (https://mcp.jsoncut.com/legacy/mcp), which still works for old integrations but receives no new authoring features. Don't configure both endpoints in one agent.
The working loop that keeps results editable
The JsonCut docs prescribe a specific agent loop, and it's worth asking for it explicitly in your prompts:
| Step | Agent action | Why it matters |
|---|---|---|
| 1 | Read the authoring guide once at the start of a task | Grounds the assistant in project rules before it touches anything |
| 2 | Use one durable project for iterative work | Tomorrow's revision continues where today's ended |
| 3 | Upload or discover owned media before referencing it | No invented assets, no broken references |
| 4 | Make bounded edits, preserving the current version | Small, reviewable changes instead of source rewrites |
| 5 | Validate, then request review frames or a contact sheet | You see the result before it's final |
| 6 | Repair only visible problems | No speculative rework of what already works |
| 7 | Render the pinned final version | The output matches exactly what you approved |
Structure follows length: for a simple short video the assistant keeps the project flat; for a long video it creates meaningful editorial scenes and works scene by scene.
This review loop is the difference between "AI generated something" and "AI produced something you can ship." Text overflow, clipped logos and awkward crops are cheap to fix on a review frame and expensive to fix after 200 renders.
Getting files in: pick the route your client supports
OAuth connects your account; it does not magically hand the assistant your local files. The docs list three transfer routes, matched to what the client can actually do:
| Client capability | Transfer route |
|---|---|
| Host supplies an authorized file reference | jsoncut_v2_import_host_file imports an HTTPS download URL (JPEG, PNG, WebP, MP4, WebM, MP3, WAV, format support depending on the host) |
| Agent can read bytes and make HTTP requests | One-use upload ticket (valid 60–900 seconds) or the mcp/upload_file.py helper, which streams bytes and verifies the SHA-256 |
| No permitted byte-transfer channel | jsoncut_v2_get_upload_handoff links to the owned Studio project; you upload under Files, then the agent lists project media |
Small programmatically encoded files (up to 3 MiB decoded) can also go through jsoncut_v2_upload_media.
From one-off renders to repeatable templates
The biggest leverage isn't the first render — it's the fiftieth. Once a layout works, the assistant can search published templates, read their typed inputs with jsoncut_v2_get_template_inputs, upload any required media and render values that satisfy the discovered contract. The docs are explicit that an agent should not infer fields from a thumbnail or title.
That's also how assistant-driven work stays honest at volume: the examples gallery shows what production-ready output looks like — every preview there comes from a real render of an editable project — and the same review-before-render loop applies whether you produce one variant or a whole campaign. If you're evaluating formats first, the video automation examples walkthrough breaks down how these projects become reusable templates.
FAQ
Do I need an API key to connect Claude or ChatGPT? No. OAuth is the recommended path for interactive clients: you sign in to JsonCut Studio and approve the connection. API keys are the alternative for clients without OAuth or for unattended automation.
Can the assistant publish my videos to social media? No. JsonCut produces rendered image and video files; distribution to platforms is separate and stays in your hands.
What happens to the project after the assistant is done? It remains a normal editable JsonCut project with version history. You can keep refining it in Studio, turn a finished version into a template, or ask the same assistant to continue later.
Sources
JsonCut documentation, "Connect an MCP client" — https://jsoncut.com/docs/mcp/ (accessed 2026-10-05)
Model Context Protocol, "What is the Model Context Protocol?" — https://modelcontextprotocol.io/docs/2026-07-28/getting-started/intro (accessed 2026-10-05)
JsonCut examples gallery — https://jsoncut.com/examples/ (accessed 2026-10-05)