The interesting part is what you do not have to decide: compose_project takes the description, including variations, and Glif resolves which models run in what order. Because it answers with a job id rather than the media, anything built on it has to poll — the finished files arrive as links to a CDN, not as bytes in the conversation, which is what keeps a batch of variations affordable to ask for.
A hosted media-generation server at https://glif.app/api/mcp: generate images, video and audio, transcribe, render HTML, search the web, run code, and chain multi-step media operations from any MCP client.
- compose_project takes a plain-language brief — including a series of variations — and Glif chooses the models and chains the steps itself
- The job returns a job_id straight away; get_job_status returns the finished media as resource_link blocks pointing at CDN URLs
- Read and upload helpers around those two, with tools/list as the authoritative set for whatever your client sees
- Streamable HTTP with OAuth 2.1 and dynamic client registration — sign in with your Glif account, no API key to place
- One-click install snippets for Claude, Claude Code, Cursor, VS Code, ChatGPT and Codex
A Glif account and a client that can reach a remote MCP server: paste https://glif.app/api/mcp as a custom connector and approve the browser sign-in on first connect. Generation spends credits from that account; the read-only tools do not. The locally run npm package is deprecated, so a config that launches it over stdio should move to the URL. MIT covers this repository, which holds the registry metadata; the hosted server itself is not open source.
One command — npx -y @glifxyz/glif-mcp-server@latest