Slop TVNewsLatest
Explainers

Affogato's MCP server makes video inside Claude and ChatGPT

Affogato says the server wraps 120-plus image, video and voice models and its studio workflows behind OAuth sign-in and per-workspace spending caps.

Illustration: Affogato's MCP server makes video inside Claude and ChatGPT
Illustration: SLOP TV News. Logos are trademarks of their owners.

Key takeaways

  • Affogato launched an MCP server at https://api.affogato.ai/mcp, announced on Y Combinator's Launch page on November 18, 2026, that connects to Claude, ChatGPT, Cursor, Claude Code and Codex.
  • Affogato says the connection reaches 120-plus image, video and voice models and studio workflows for talking heads, lipsync, voiceovers, headshots and virtual try-on.
  • Sign-in is OAuth with no API key to paste, and the agent spends only a workspace the user approves, with estimate_cost quotes, a confirmation threshold, a daily cap and full refunds on failed or cancelled generations.
  • The launch offer code MCP10 takes 10 percent off the first month on any Affogato plan until January 17, 2027.

Affogato turned its AI studio into an MCP server, so an assistant can make the images, video and voice instead of only describing them.

The company announced the connection on Y Combinator's Launch page on November 18, 2026, with a setup guide at affogato.ai/mcp. Add the address https://api.affogato.ai/mcp to Claude, ChatGPT, Cursor, Claude Code or Codex, sign in once, and the assistant can call Affogato's models and studio workflows from the same chat where the script was written.

MCP, the Model Context Protocol, is the open standard that lets an AI assistant use outside tools. Affogato's version puts a full media studio behind that door.

What is Affogato's MCP server?

It is one connection to Affogato's catalog and its guided workflows, reachable from inside an AI assistant. The launch post says the server reaches "120+ AI models" for image, video and voice, drawn from the same catalog as the app. The MCP page puts the count at 123 image, video and audio models, and the registry listing for the server, ai.affogato/affogato, repeats the 120-plus figure in its own description.

Affogato AI is the studio formerly known as RenderNet, backed by Y Combinator, selling images, video, voice and "AI humans that don't look AI" from one credit balance. Its homepage lists eight studios, including Image, Video and Talking Head studios, plus a Video Editor. The MCP server exposes the studio workflows as single tools, so a user asks for the outcome and the agent picks the steps.

What can the assistant make?

Affogato says the server wraps five studio workflows as single tools: talking heads, lipsync, voiceovers, headshots and virtual try-on. The launch post's example prompts are the practical version: a professional headshot from a selfie, a script voiced and lipsynced onto a clip, a hoodie put on a model in a street scene, a 15-second talking-head video of the user reading a script.

The MCP page lists the calls behind those workflows, including generate, lipsync, talking_head_keyframe, talking_head_video and cast_swap. Files stay simple. When a request needs a photo, a clip or a voice track from the user's device, Affogato says the assistant sends an upload link and carries on once the file arrives.

How do you connect it?

With OAuth, and no API key to paste. The first request opens an Affogato sign-in page, where the user picks the workspace whose credits the agent may spend and approves it. The consent page lists three permissions: read (list models and read generations, files and account), generate (spend the connected workspace's credits) and files (add images, video and audio as inputs). A user can disconnect an agent from the Developers page of the app.

For coding agents, Affogato gives the command line: claude mcp add --transport http affogato https://api.affogato.ai/mcp. It also publishes Agent Skills, which teach an agent the studio workflows, under an MIT license at github.com/affogato-ai/skills, plus a Claude Code plugin that installs the skills and the server together.

How does it keep an agent from overspending?

Affogato built the spending rules for an unattended agent, and they run per workspace. The estimate_cost tool quotes a request in credits and never charges. A generation above the workspace's confirmation threshold does not start until the agent checks with the user. A daily credit cap stops runaway loops and resets every day. Failed or cancelled generations are charged up front and refunded in full.

The agent spends the same credits the team spends in the app and the API, one wallet rather than a second budget. Affogato says the connected workspace needs a paid plan or purchased credits, and that switch_workspace moves a session to another workspace the user belongs to.

What does the talking-head workflow produce?

Talking Head Studio turns a script into a multi-segment presenter episode, and Affogato lists two engines for it: Kling AI Avatar v2 or ByteDance OmniHuman 1.5. The studio's own page says each paragraph becomes a segment of up to about 45 seconds, stitched into an episode of up to about 570 seconds, just under ten minutes, in 1920x1080 or 1080x1920.

The host is any saved character, including one made from the user's photo, on one of ten curated sets or the user's own. The voice comes from a 58-voice library, a custom voice, or the user's own recordings. A table read voices the whole episode for a few credits before any video is rendered, and the page says editing one line re-renders only that segment. Kling AI Avatar v2 runs are listed at 102 credits in the catalog and OmniHuman 1.5 at 290.

What does it cost?

Plans start at $24 a month for 2,000 credits, and every paid plan carries a commercial license with no watermarks. Affogato's pricing page lists Ultra at $69 a month for 7,000 credits, a Pro Max tier at 35,000 credits, and Team from $99 a month with 25 seats and a shared pool. Unused credits roll over up to a cap that varies by plan.

The launch offer code MCP10 takes 10 percent off the first month on any plan, and the post dates it to January 17, 2027. There is no free plan.

What is still unknown?

Affogato's own pages describe the server, and independent measurement is only starting. A directory that tests MCP servers, INITE Atlas, read the endpoint on November 17, 2026 and found it answering and asking for OAuth sign-in, with the tool list visible only to signed-in users and a median first answer of 119 milliseconds. It had run two of the seven daily checks it uses before it prints a reliability score, so it does not yet grade the server.

The official MCP registry lists the server as ai.affogato/affogato, version 1.0.0, published November 17, 2026. Affogato's capability claims are its own, made on launch day, and none of them have been tested here.

The MCP server is live for anyone with a paid Affogato plan; the setup guide is at affogato.ai/mcp.

Sources

  1. ycombinator.com - maker's launch post: MCP server, 120+ models, OAuth, spend rails, MCP10 offer, founders
  2. affogato.ai - setup guide: 123 models, three permissions, estimate_cost, daily cap, tool list, Agent Skills
  3. registry.modelcontextprotocol.io - official MCP registry: name ai.affogato/affogato, version 1.0.0, published 2026-11-17
  4. affogato.ai - Talking Head Studio: segment length, episode length, engines, credit costs
  5. affogato.ai - plan prices, monthly credits, rollover caps, commercial license
  6. affogato.ai - 123 models from 35 providers, presenter engines
  7. inite.ai - independent endpoint measurement: OAuth, tool list hidden, 119 ms median answer, read 2026-11-17