KinoPipe

KinoPipe is an FFmpeg-as-a-service platform built specifically for AI agents and developer backends. Created and run by independent developer Nicolas Coutureau, it exposes 48 typed video editing operations — trim, resize, compress, captions, scene splitting, and more — through a remote MCP server and a REST API, so agents never need to touch a shell or bundle an FFmpeg binary.

What It Is

KinoPipe sits between an AI agent (or any backend) and the complexity of video processing. Instead of generating FFmpeg shell commands, an agent sends a validated JSON request to a typed endpoint and receives a finished file URL in return. Every operation is a stable, bounded contract: named inputs, typed options, and a predictable response shape. The platform runs hosted FFmpeg 9 workers with GPU encoders, handling the rendering infrastructure entirely on its side.

How the Agent Workflow Works

The typical flow has three steps: describe the edit (send a video URL and typed options), wait for rendering (KinoPipe processes in a single pass on its workers), and retrieve the file (a download URL or a signed webhook fires when the job is done). Agents connect via the remote MCP server at kinopipe.com/mcp using OAuth 2.1, which means no API key to paste into a prompt. REST API callers use bearer keys and can attach idempotency keys for safe retries. The Jobs API accepts multi-operation pipelines — trim then resize, for example — processed in order in one encode.

The 48 Typed Tools

Each tool is a focused endpoint with its own page, stable slug, and narrow option set. The available prebuilt tools include:

• Trim, resize, compress, crop — core editing primitives
• Add subtitles — burn SRT or WebVTT into video
• Merge videos — normalize and join two to six clips
• Split by scenes / segment duration — up to 50 output files
• Convert to GIF — palette-optimized animated GIF
• Add music, watermark, picture-in-picture — compositing operations
• Extract audio, remove audio, generate thumbnails — utility tools

The Jobs API also supports composing these as operations in a single request for advanced pipelines.

MCP and Agent Client Support

KinoPipe publishes a remote MCP server with OAuth 2.1 and Streamable HTTP transport. The About page and homepage list supported clients: Claude, Claude Code, Cursor, Windsurf, GitHub Copilot, Codex, ChatGPT, and n8n (via a community node). One command or config entry adds every tool to the agent's discovery context. The same operations are available over plain REST for backends that don't use MCP.

Design Philosophy and Tradeoffs

The About page is explicit about the product's constraints and values. KinoPipe never accepts shell commands — only bounded, validated operations — which limits surface area and reduces the risk of injection or unexpected behavior in agent loops. Files are not retained: direct uploads are deleted after 24 hours, outputs are served through expiring signed URLs for up to 12 months, and media is never used to train models. Failed jobs automatically refund their credits. The product is built and supported by a single founder, who the About page notes answers support chat directly.

Current Status

KinoPipe is actively available with a free tier requiring no credit card. The homepage cites a 325-run benchmark covering cold, warm, and burst phases, with a reported p50 render time of 8.2 seconds for a 1080p edit and a 100% success rate across those benchmark runs. The platform is positioned as production-ready for agent workflows, with an OpenAPI 3.1 spec published at /openapi.json for client generation.