Skip to content
RiffAds
Log in

MCP server

The MCP server for Claude video generation.

The RiffAds MCP server brings video generation to Claude, ChatGPT, Cursor and other MCP clients, so your AI agent can make UGC video ads with AI actors. It prices every job before it spends and stops at limits a person sets.

Open the studioConnect your agent

What it is

What is an MCP server for video?

The RiffAds MCP server is a remote MCP server for video ads, at mcp.riffads.com. Connect it in Claude, ChatGPT or Cursor and your AI agent gets video generation: it prices a job, makes a talking actor ad, an image or a video, and hands back a download link. It spends the same credits as the studio.

Server URL

https://mcp.riffads.com
Transport
Streamable HTTP. It is a remote server, so there is nothing to run on your computer.
Sign-in
OAuth, in your browser. There is no API key to copy. Your app finds the sign-in page through the RFC 9728 metadata the MCP spec asks for.
Access
You choose when you approve it: read only, which can browse and price but never spend a credit, or generate and spend credits.
Workspace
In more than one workspace? You pick which one it spends from, and you can disconnect it on the Connections page.

Works with

Video generation in
Claude, ChatGPT or Cursor.

Each of these apps connects with the URL above. The connect page has the exact button or command for each one.

Step by step for each app

Any other MCP client that supports remote servers and OAuth sign-in can use the same URL.

Tools

Every tool
your agent gets.

Grouped by the job in hand. Only submit_generation, generate_talking_actor and run_workflow spend credits. Everything else is free to call.

The RiffAds MCP tools, grouped by job
JobToolsWhat they do
To find and price
  • list_capabilities
  • get_capability_schema
  • list_actors
  • list_voices
  • list_presets
  • estimate_generation
  • get_credit_balance
See what RiffAds can make, the settings each job takes, the actors, voices and presets, and your credit balance and spend limits. estimate_generation prices a job before anything is spent.
To make
  • submit_generation
  • generate_talking_actor
Start a job. submit_generation runs any capability by its id, and generate_talking_actor is the same call, named for talking actor ads. Both need max_credits.
To follow a job
  • wait_for_generation
  • get_generation
  • get_batch
Wait on the server, up to 20 seconds per call, until a job is done, then read the result. get_batch reads a set of variants together.
To read history and the library
  • list_generations
  • search_library
  • get_asset
Look up past jobs and your library of uploads and outputs, and get a fresh link to a file.
To bring files in
  • create_upload
  • finalize_upload
  • import_media_from_url
  • open_upload_box
  • send_file_bytes
Upload a photo, video or audio file, or bring one in from a public link. Each file is checked before a job can use it, and images also get a safety check. Neither step costs credits.
For workflows
  • list_templates
  • run_workflow
  • get_workflow_run
  • list_workflow_runs
List the ready-made workflow templates, run one in a single call with one max_credits for every step, and follow its runs.
To check the connection
  • riffads_ping
Check that the connection works, which workspace and plan it uses, and whether it may spend.

Start to finish

How an AI agent
makes a video ad.

The steps

  1. 01

    Find an actor and a voice

    list_actors and list_voices return the AI actors and voices your workspace can use.

  2. 02

    Price the job

    estimate_generation is free. It returns the credits and max_credits_needed: the price plus a small margin that is held while the job runs.

  3. 03

    Make the voice-over

    Call submit_generation with the tts capability, the script, the actor, the voice and max_credits. Some actors speak for themselves and skip this step.

  4. 04

    Make the video

    Call generate_talking_actor with the finished voice-over and max_credits. A job over max_credits or over a limit is refused before it starts, never trimmed.

  5. 05

    Wait for it

    wait_for_generation waits on the server for up to 20 seconds per call. The agent calls it again until the job is done.

  6. 06

    Get the file

    The answer carries a download link that works for 10 minutes. get_asset gets a fresh one later.

mcp.riffads.com

Make a 15 second ad for my earbuds. Keep each job under 600 credits.

list_actorsactor picked
list_voicesvoice picked
estimate_generationvoice-over priced
submit_generationvoice-over started
wait_for_generationvoice-over ready
estimate_generationvideo 420 credits, max_credits_needed 462
generate_talking_actorvideo started, max_credits 462
wait_for_generationstill_running, call again
wait_for_generationdone

ad.mp4 ready. Download link valid for 10 minutes.

A sample run: the tools an agent calls to make one talking actor ad.

For this 15 second ad the video is about 420 credits and is held as 462, under the default cap of 600 credits a job. The voice-over is priced and held the same way, as its own job.

Spend limits

Spend limits
a person sets.

Your agent names the most it may spend on every paid call. A person sets the limits around that number, and the agent cannot change them.

Set your agent limits
The four limits on what an agent may spend
LimitSet byDefaultMeasured over
max_credits on the requestYour agent, on every paid callRequired, no defaultOne job
Per job capA workspace owner or admin600 creditsOne job
API key budgetA workspace owner or adminNo limitRolling 24 hours, per REST API key. An MCP connection has no key, so it does not apply here.
Daily agent capA workspace owner or admin2,000 creditsRolling 24 hours, all agent spending in the workspace

  • The smallest limit wins

    A job is checked against the lowest limit that applies, and every refusal names the limit that stopped it, in limit.bound_by.

  • Checked before a job starts

    Limits decide whether a job may start. A job that has started is never stopped halfway, because the work is already being done.

  • Zero means no spending

    Set a cap to 0 and agents in that workspace cannot spend at all. Or connect an agent read only, so it can browse and price but never spend.

Never

What your agent
can never do.

  • Post anywhere

    It never posts to Meta, TikTok, YouTube, X or any other social platform.

  • Return anything but results

    It returns a download link to each finished file, or for an analysis, the JSON breakdown itself. Where the ad goes next is your call.

  • Spend past your limits

    By default an agent may spend 600 credits a job and 2,000 credits a day. A workspace owner or admin can lower both, or set them to zero.

About MCP

MCP
in numbers.

MCP, the Model Context Protocol, is an open standard for connecting AI assistants to tools and data. Each number below links to where it was published, with its date.

FAQ

MCP server, answered.

Still unsure? Book a call

An MCP server gives an AI assistant tools through the Model Context Protocol, an open standard Anthropic released in November 2024. The assistant connects to the server, reads its list of tools and calls one when a task needs it. The RiffAds MCP server gives it tools to price and make video ads.

Claude, Claude Code, ChatGPT, Cursor and VS Code can all connect with one URL, and so can other MCP clients that support remote servers with OAuth sign-in. Claude allows custom connectors on every plan, including Free. ChatGPT adds them in developer mode, on Pro, Plus, Business, Enterprise and Education accounts on the web.

No. API, MCP, Agent Skills and CLI come with every plan. Your agent spends the same credits as the studio, at the same prices, and pricing a job before it runs is free.

No. Every paid call must set max_credits, the most it may spend, and a job that costs more is refused, never trimmed. By default agents are also capped at 600 credits a job and 2,000 credits a day per workspace. An owner or admin can lower both or set them to zero, and every refusal names the limit that stopped it.

No. RiffAds never posts to Meta, TikTok, YouTube, X or any other platform, and it never connects to an ad account. Your agent gets a download link to the finished file, and you decide where the ad runs.

Not on its own. Claude has no video model, so it cannot render a talking actor ad or a video clip by itself. Add the RiffAds MCP server as a custom connector and Claude gets video generation tools: it prices the job, makes the ad, image or clip, and hands back a download link, spending your workspace's credits within limits a person sets.

No. Add the URL as a connector in Claude, ChatGPT or Cursor, approve it in your browser, then ask for an ad in plain words. Your agent picks the actor, checks the price and brings back the file. You only need code to call the REST API from your own software.

Stop waiting on creators. Ship the ad today.

Book a call for a walkthrough, or log in and make your first ad.

Book a callLog in
New here? Try Growth for $1