Connect your AI

Add Typecut to Claude, ChatGPT, Claude Code, Cursor or any MCP client, so it can plan, voice, build, check and share videos for you.

The Typecut connector lets an AI agent do what you do in the app. It plans a project, voices it, builds and checks scenes, and hands you a link, all from inside a conversation. It works on every plan, including Free, and spends the same credits as the app. There is no separate API price.

The address

MCPhttps://app.typecut.ai/api/v1/mcp

This is the one most people want. It speaks Streamable HTTP and keeps no session between calls. When your client connects it sees a short list of tools grouped by the step of the work, and every other Typecut tool is still callable by its name (see the tool reference).

Two more doors exist for special cases:

Address	What it lists	Use it when
`https://app.typecut.ai/api/v1/mcp`	A short list by step, plus `tools` to find the rest	Almost always
`https://app.typecut.ai/api/v1/mcp/studio`	The tools for writing each scene as HTML	Your agent authors scenes itself (Write your own HTML)
`https://app.typecut.ai/api/v1/mcp/house`	The old full menu, every tool listed	A client that cannot call unlisted tools. Being retired

All three take the same sign-in and reach the same projects.

Connect your client

You can sign in two ways: OAuth (paste the address and approve in the browser) or an API key. OAuth is the one to use for yourself.

In Claude, open Settings → Connectors and choose Add custom connector.
Paste https://app.typecut.ai/api/v1/mcp.
Claude opens the Typecut consent screen. Approve it.

No API key is involved.

Add Typecut as a custom connector in ChatGPT and paste https://app.typecut.ai/api/v1/mcp. ChatGPT finds the sign-in on its own and opens the Typecut consent screen. Where the connector setting lives depends on your ChatGPT plan, so check ChatGPT's own connector settings.

Scene and project players appear in the chat where ChatGPT supports them.

Terminal

claude mcp add --transport http typecut https://app.typecut.ai/api/v1/mcp

Then run /mcp inside Claude Code and choose Authenticate. Or skip OAuth and send a key:

Terminal

claude mcp add --transport http typecut https://app.typecut.ai/api/v1/mcp \
  --header "Authorization: Bearer $TYPECUT_API_KEY"

Add this to your MCP config (~/.cursor/mcp.json). Windsurf and other config-file clients use the same shape.

mcp.json

{
  "mcpServers": {
    "typecut": {
      "url": "https://app.typecut.ai/api/v1/mcp"
    }
  }
}

Cursor asks you to sign in the first time it connects. To use a key instead, add "headers": { "Authorization": "Bearer tc_live_…" }.

Any client that supports remote MCP over HTTP works, including Codex CLI. Give it the address. If it supports OAuth it finds the sign-in by itself. Otherwise send Authorization: Bearer tc_live_… with every request. Check your client's own documentation for where the setting lives.

OAuth or API key?

	OAuth	API key
Setup	Paste the address and approve	Create a key in Settings → API keys and paste it into config
Plan	Any, including Free	Creator and up
Lifetime	Access lasts an hour and renews by itself; the sign-in lasts up to 90 days	Until it expires or you revoke it (new keys default to 90 days)
Best for	People	Scripts, CI and shared service accounts

Without a credential, the server answers 401 and points your client at the sign-in, which is what turns "unauthorized" into a Connect button. Disconnect an AI app under Settings → AI apps (MCP), and revoke a key under Settings → API keys.

How an agent works

The tools on the main list follow the steps of making a video. A short piece takes about a dozen calls.

Orient. get_account gives the plan, balance and workspaces. list_catalog lists valid voices, themes, models and prices.

Plan. create_project with a plan lays out chapters, blocks and scenes exactly, in one free call. suggest_plan drafts one from a topic or script first. list_skills and get_skill pick a visual style before any scene is written. apply_changes reshapes the structure later as one undoable edit.

Sound. voice_project voices the narration with word timings (price it with dry_run: true first). speak voices single lines and dialogue. set_music puts a bed on the beat grid. get_timeline shows every track on one clock.

Media. find_media searches your shelf, your library, other projects and stock in one call. generate_image and generate_video make what is missing. decompose_image takes a picture apart into layers for parallax.

Build. get_scene_context returns what a scene must fit: its beats, the words on the scene's clock, bars and the look. Typecut can design the scene (generate_scene), or the agent writes it and put_scenes lands it through the real checks. Landing HTML an agent wrote is open to pilot accounts for now; every account can check it with dry_run.

Look. look returns frames, a motion strip and sync findings for any stretch of the film. It is free. analyze_video reads a reference video shot by shot.

Deliver. While it works, the agent hands you the editor link, https://app.typecut.ai/project/<project_id>. share makes a public link only when you ask for one, and render makes the MP4 when the piece is finished or you ask.

Slow tools run in the background: the call returns a job, and jobs (or get_status for voice, generation and analysis ids) waits up to about 50 seconds for it in one call. comments reads and answers review comments people leave in the app.

Pricing and credits

An agent pays what you would pay.

Tools that spend credits say SPENDS CREDITS in their description, and are marked in the tool reference. Everything else is free.
Agents can price first. Many spending tools take dry_run: true. quote prices a list of operations and returns a ceiling and a quote_id. get_price_book lists every price.
hold_credits reserves a quote's ceiling so parallel work cannot overspend. create_budget gives an agent you spawn its own key with a hard cap.
get_account before and after is the receipt.

Good agents show you the number before they spend. You can ask for that in your first message.

How it behaves

Everything is a project. An agent makes the same projects the app does. You can open, change and export them yourself.
Edits do not overwrite you. Edit tools carry the version they read. If you changed the same scene in the app, the agent gets a conflict instead of silently replacing your work.
Errors are readable. A bad value comes back with the valid ones, so the agent can correct itself.
Old names keep working. Tools that were renamed or moved off the main list (for example put_scene_composition, preview_composition, review_motion) still answer by their old names.

Prefer plain HTTP?

Each tool forwards to a /v1 route with your credential. See API vs MCP vs app and the API reference.