FrameThrower

What the FrameThrower MCP server does

FrameThrower MCP server exposes a film-still reference library through a remote MCP endpoint. The library contains frames from 5,489 films, indexed by visual properties such as lighting, lens character, shot size, colour, and mood. Search results identify the film, director, and cinematographer, and include a thumbnail URL, colour palettes, metadata, and a link back to FrameThrower.

The service is intended for discovery and visual study rather than image delivery. It returns references and links, not raw image bytes. The frames are shown at reduced resolution, and rights remain with their respective owners.

How it works

Connect an MCP-compatible client to https://framethrower.ai/api/mcp and complete the OAuth 2.1 sign-in flow. After connecting, the server supplies usage instructions that guide the agent through an iterative reference workflow:

1. Search for a visual idea.
2. Present a small selection with film information and links.
3. Refine the selection using feedback such as “more like this,” “colder,” or “tighter.”
4. Repeat until the desired reference is found.

The FrameThrower MCP server keeps a link with each result because chat clients may not reliably render inline images, while the FrameThrower gallery provides the full reference context.

Setup and configuration

No local installation or long-running process is required for clients that support the remote MCP URL. In Claude, add a custom connector and paste the endpoint, then sign in when prompted. In Cursor, add the endpoint to ~/.cursor/mcp.json or use the provided installation link.

Clients that only support stdio can connect through mcp-remote. The bridge runs with npx and forwards requests to the hosted endpoint while handling OAuth. The repository contains per-client setup instructions, tool signatures, manifests, and examples; it does not contain a server that you run locally.

There is no API key to configure. Authentication uses the signed-in FrameThrower account. New accounts receive $2 in credits without entering a card. After that, credits cost $1 per 1,000, and each tool call uses 2 credits. Credits do not expire, and the service has no rate limits.

Tools and capabilities

The FrameThrower MCP server provides four tools:

• search_frames accepts a required natural-language query and an optional result limit from 1 to 40. It suits concepts such as a neon-lit rainy street at night.
• find_by_craft filters by lens, shot size, visual style, setting, time of day, director, and year range. All filters are optional and can be combined.
• find_similar takes a frame ID from an earlier result and returns related frames for refinement.
• get_frame_details takes a frame ID and returns detailed information for that frame.

Each tool call costs 2 credits. Results include metadata, thumbnails, colour palettes, and deep links, but not image data. The same account can also be used with the service’s REST API and npm SDK, although those are separate interfaces from MCP.

Limitations and notes

This is a hosted service, so it cannot be self-hosted from the repository. The repository acts as a connector and documentation package. A client that only supports stdio needs the mcp-remote bridge.

Use the results as cinematography references for commentary, education, and study. They are not a licensing or image-download pipeline. Running out of credits produces a response containing the remaining balance and top-up location rather than failing silently.