Clipwright

@clipwright/mcp-server

Your coding agent writes the script; Clipwright renders a short vertical video of an actor speaking it, 1080×1920 by default. It also makes faceless videos of a length you pick, 30 to 90 seconds, from a script or a short brief: narration over an opening animated clip and image scenes, with no on-camera presenter. The agent asks for a free quote before it spends anything, and whatever the API cannot honor comes back in warnings[] instead of being dropped silently.

The agent gets a run_id at once and polls until the video URL is ready. Clipwright returns the file; it does not publish anywhere. Quotes, voices, actors and the account balance are free. Renders spend credits: an account opened by signing up starts with 420 trial credits, enough for one 90-second faceless video. Buy more at clipwright.io/pricing.

Install

There are two ways to connect: the hosted server at https://mcp.clipwright.io/mcp, or this package run locally over stdio. Both serve the same tools.

Hosted server, no key

Add the URL to your client. On first use the client opens a browser, you sign in to Clipwright and approve the app. Claude Code and the device-code flow below were verified live on 2026-10-18; the other commands come from each client's own documentation and have not been tried with Clipwright yet.

• Claude Code (verified live): claude mcp add --transport http clipwright https://mcp.clipwright.io/mcp, then run /mcp in a session and sign in.
• Claude.ai and Claude Desktop: Customize → Connectors → Add custom connector, then paste the URL.
• ChatGPT: add the URL as the remote MCP server of a plugin, as OpenAI's guide "Connect and test your plugin" describes; ChatGPT signs in with OAuth.
• Cursor: in ~/.cursor/mcp.json, {"mcpServers": {"clipwright": {"url": "https://mcp.clipwright.io/mcp"}}}.
• VS Code: in .mcp.json at the workspace root, {"mcpServers": {"clipwright": {"type": "http", "url": "https://mcp.clipwright.io/mcp"}}}.
• Gemini CLI: gemini mcp add --transport http clipwright https://mcp.clipwright.io/mcp, then /mcp auth clipwright.
• Codex: codex mcp add clipwright --url https://mcp.clipwright.io/mcp, then codex mcp login clipwright.

The approval page lists the permissions the app asks for. There are two:

• mcp:read — the tools marked read-only: account, runs, shots, voices, actors, workflows and quotes.
• mcp:write — every other tool. Some start paid runs and spend credits; the rest change the account (actors, uploads, workflow plans).

A connection without mcp:write gets 403 insufficient_scope from those tools, and nothing is charged. Only an admitted account can approve.

To disconnect an app, open app.clipwright.io/api-keys, find it under Connected apps and revoke it. Its next request gets 401. A connection lasts 90 days from approval; after that the app asks you to sign in again.

On the hosted server upload_image cannot read a file on your machine. Pass a public https URL as upload_image({url}) and the server downloads it, or call create_upload and PUT the file to the returned put_url with curl. Either way you get an upload_id for first_frame, last_frame and references.

An agent with no browser signs in with a device code and the first-party client clipwright-agent (verified live 2026-10-18):

curl -s https://mcp.clipwright.io/device/auth -d client_id=clipwright-agent \

--data-urlencode "scope=mcp:read mcp:write offline_access" \ --data-urlencode "resource=https://mcp.clipwright.io/mcp"

Open verification_uri (https://mcp.clipwright.io/device) on any device, enter user_code, sign in and approve; the approval page repeats the code. Meanwhile poll POST https://mcp.clipwright.io/token with grant_type=urn:ietf:params:oauth:grant-type:device_code, the device_code and client_id=clipwright-agent, no more often than every 5 seconds: a faster poll gets slow_down. Send the access_token as Authorization: Bearer to /mcp; it lasts an hour. Renew it with grant_type=refresh_token and keep the newest refresh token each time. The device flow is open to this client only.

Hosted server with a cw_* key

For CI and scripts, send a key instead of signing in. A key holds both permissions.

claude mcp add --transport http clipwright https://mcp.clipwright.io/mcp --header "Authorization: Bearer cw_..."

n8n

n8n's built-in MCP Client node calls one tool per node:

• Server Transport: HTTP Streamable; MCP Endpoint URL: https://mcp.clipwright.io/mcp.
• Authentication: Bearer Auth, with a credential whose token is your cw_... key.
• Tool: pick from the list, for example quote_ugc; Input Mode: JSON.

Verified live on 2026-10-20 with self-hosted n8n 2.43.8: get_account, quote_ugc and get_run returned through the node. A workflow to import, n8n/clipwright-ugc.json in seocombat/clipwright-mcp, quotes a script, stops when the quote is above max_credits, starts make_ugc and polls get_run every 15 seconds. The whole workflow ran on n8n Cloud on 2026-10-20 and returned a 1080×1920 MP4 in about three minutes. n8n's OAuth credential has not been run against Clipwright yet.

Local server over stdio

claude mcp add clipwright -e CLIPWRIGHT_API_KEY=cw_... -- npx -y -p @clipwright/mcp-server clipwright-mcp

Put your own key in place of cw_.... The command is one line and gives the shell nothing to expand, so bash, zsh and PowerShell run it alike. It needs Node.js 20 or newer.

Any MCP host works — the local server speaks JSON-RPC over stdio. Issue keys at app.clipwright.io/api-keys.

Environment

These variables configure the local server.

Recipes

Each recipe is a prompt we gave an agent with this server connected, the call the agent sent, and what the video cost. The videos play at clipwright.io/mcp.

1. A creator reel with on-screen graphics

One presenter on camera, with graphics over her that react to what she says: the format of talking-head reels that sell a product. One prompt, no editor.

The prompt, given verbatim, with our actor id, screenshot link and site named alongside it:

Make a 30-second vertical reel for my AI receptionist with my presenter actor and the voice Hope. Script: "Every missed call is a lost deal. Your receptionist sleeps, takes breaks, and sometimes gets annoyed. This one doesn't. It answers calls all day, works around the clock, and never gets annoyed. Here's a real one: a customer calls at two in the morning to book a cleaning, and it handles everything, the time, the price, the confirmation text. Small businesses miss a third of their calls. Not anymore. Set it up in ten minutes and stop losing deals while you sleep."

While she lists what it does, show a checklist with three switches ("answers calls" and "works 24/7" off, "gets annoyed" on) and have a cursor flip each one as she says it. When she gets to the 2 a.m. customer, show an incoming call from +1 (305) 555-0147 on the right and answer it on "handles". Put my dashboard screenshot in the top right corner while she says "Small businesses miss a third of their calls." End on a card with my site and "create your assistant". Caption only the key lines, in the highlighter style, and zoom in a little on "Not anymore." Quote it before you render.

What the agent sent: one make_ugc quote with templates (checklist, phone_call, end_card), one overlays picture, caption_mode: "highlights", caption_style: "highlighter" and a punch_in on one emphasis line. Click, switch and ring sounds come with the templates by default. We checked the prompt with a fresh agent on this server on 2026-10-24; the video was rendered from the same call.

The quote named the price before anything was spent; list_prices names every rate. Watch it.

Tools

• list_workflows — read workflow methods and current availability. Free.
• get_workflow — read one method's required inputs, script rules, evidence and limitations. Free.
• prepare_workflow_plan — create with a complete document, or save with both plan_id and base_version. Free; starts no run.
• plan_reference_remake — the server compiles a reference ad's profile and your exact script into a new plan; its warnings name what the plan does not reproduce. Free; starts no run and generates nothing.
• get_workflow_plan — read the current owned plan, revision and readiness. Free.
• approve_workflow_plan — explicitly approve a chosen revision and content digest. Free; starts no run.
• quote_workflow_plan — quote the approved revision without spending credits.
• run_workflow_plan — start a paid workflow using its exact quote fields and an explicit idempotency key. Returns its receipt and run id immediately; poll get_run. Recover an uncertain purchase with the original request and key.
• list_voices — voices for voice: the presets, the thirty voices of the speech model gemini-3.8-flash-tts and the catalog for make_ugc, with filters by language, gender, age, use case and model, and a short audio sample where the voice has one. The seventeen narration voices of make_faceless are listed with skill: "make_faceless". Free, spends no credits.
• list_template_sounds — the sounds make_ugc templates play, by beat: where each beat plays, its default and every name template_sounds takes. Free.
• list_actors — ready-made faces with stable ids and verified formats. Free.
• quote_ugc — estimated duration and price, plus a warning when make_ugc with the same input would be refused for money or access. Free.
• get_account — balance, debt and holds of the account. Free.
• list_prices — what every tool costs in credits: a number per unit, or a formula where the price depends on the input; free tools are listed at 0. Free.
• send_feedback — tell Clipwright what is missing, confusing, wrong or slow, optionally about a run; we read every one. Free.
• make_ugc — renders a clip; this is the call that costs money. Returns a run_id immediately instead of blocking, because a render takes minutes.
• get_run — status of a run. Free; poll it every few seconds.
• get_run_shots — the shots of a finished video: the span, kind, description and preview link of each, with the narration words and their timing. Free and read-only. A run with no shots to show answers with the reason instead.
• quote_revision — price of a revision of a faceless video: what it would change and what it costs. Free.
• revise_run — starts a revision: a new run that renders the video again with the shots you change; this costs money. Pass changes (a shot and what to generate for it) or the whole edited document. A revision that delivers is charged its quoted total; one that fails is charged nothing. An identical request returns the same run; attempt starts a second one on purpose. Returns a run_id immediately; poll get_run until succeeded.
• quote_candidate — price of a shot candidate of a faceless video: one pict