Cloudflare releases Streamline for custom video processing pipelines

Streamline opens Cloudflare video customization

Cloudflare has released Streamline as a developer playground and open-source system for creating custom video experiences on its Developer Platform. It shows how Workers, Containers and Durable Objects can be combined with media protocols to modify video and publish the result as a livestream or new hosted video.

Examples include transparent annotation overlays that can change during processing, subtitles burned into hosted video, filters, picture-in-picture, and input from webcams or factory cameras. Cloudflare has also released the media engine and an example Worker application with an Astro web frontend as open source, along with a public hosted playground.

How the Streamline architecture works

A Streamline deployment has two main components. The Media Engine runs in a Container and handles all media input, output and processing. Its controller is an HTTP server written in Go, while an FFmpeg processor performs the media work. The current processor is an internal implementation detail rather than part of the user-facing API.

The controlling Application uses Workers and can be a browser interface, an agent or an embedded system. A Durable Object orchestrates sessions, the Container lifecycle and the preview relay. The engine can pull RTMPS from one Stream Live input and publish to another, ingest Cloudflare Stream HLS manifests and segments, accept supplied video, or send preview fragments through a WebSocket relay. Processing continues if the controlling Worker disconnects. A maximum session duration ensures shutdown, and activity is renewed until that limit is reached.

Configuring and running media pipelines

The session API lets an application create or resume a session, start a pipeline from a JSON configuration, ingest video chunks, update a PNG annotation, receive metrics and stop processing. Configurations define inputs, operations and outputs. Inputs can include an RTMP broadcast, HLS-hosted video or webcam media. Outputs can use RTMP or WebSocket preview delivery through the Durable Object relay.

One example applies a transparent image overlay to a live broadcast and sends the modified result to another Stream Live input. Another ingests a Cloudflare Stream video, reads embedded closed captions, renders them onto the picture and sends the result through RTMP. Streamline currently supports filter, overlay, subtitle and encode operations. The engine applies them in a fixed order, so the order of entries in the configuration array does not change execution.

Security, limits and possible extensions

The owner deployment is private through Workers Access. The Worker verifies the Access session, binds an active session to the verified principal and allows only one session at a time. Stream Live Input keys are held in Worker secrets or write-only shared overrides in Durable Object storage, and are not returned by the settings API or placed in browser storage. Applications refer to named input and output profiles, which the Worker resolves before contacting the Container.

Preview publishing uses a Cloudflare Access service token for the container workload and a random per-session capability for the active relay. The public playground applies one container identity per verified user, one active session per user, global admission control, concurrency and media limits, and prevents one user from replacing another user's session.

Streamline currently uses Container CPU for processing, which Cloudflare identifies as a bottleneck at higher qualities or frame rates. Possible extensions include computer vision pipelines, hardware-accelerated processing, WebRTC and MoQ experiences, and native video encoding and decoding in Workers.