Automated After Effects video rendering via API

Challenge

• Short videos with an AI presenter, B-roll, music and transitions were made by hand: a designer opened the project, swapped footage, fixed timing and exported each video.
• Rendering runs only on licensed Windows machines that have to be reached over the public internet, sometimes through proxies.
• Designers keep editing compositions, so the backend cannot hard-code layer structures.
• Renders take minutes and can fail during download, render or upload, and a failure must not leave a machine locked or a job stuck.
• Some inputs come from external AI services that are asynchronous and slow.

Solution

A template is a typed contract: an .aep project, an ExtendScript file and a list of layers (video, photo, audio or data), each marked required or optional. A job fills that contract with footage and per-layer settings, and the API refuses to render until every required layer is filled. The render itself is a durable Inngest workflow made of checkpointed steps: allocate a node, stage the assets, render, upload, release the node. Inside After Effects, the template script reads the job parameters and builds the compositions, so one project file produces many different videos. Optionally, one request makes a whole video: an LLM writes the content plan and B-roll ideas, an avatar service records the narration and a text-to-video model generates clips in parallel.

What the system does

Templates and jobs API

Create and edit templates and layers, upload files straight to storage through presigned URLs, and set per-layer timing, transform, opacity, audio levels and transitions on each job.

One-call rendering

POST /render/{job_id} queues the job, and the result comes back as a download link that stays valid for 7 days.

Three-phase progress

Asset download, render and result upload each report their own percentage, taken live from the node.

Render node management

Register and remove nodes through the API or an admin CLI. SSH access is checked when a node is added, and a monitor probes every node once a minute.

End-to-end video generation

Content plan, avatar video, B-roll ideas, parallel B-roll generation, layer assembly, render and upload run as one workflow.

Web app

A React app that also runs as a Telegram Mini App: templates, jobs, avatar and B-roll generation, a timeline-style compose view, rendering and playback.

Engineering decisions

• Every render stage is a separate checkpointed Inngest step. A failed step is retried on its own and completed steps are not re-run. Errors that a retry cannot fix, such as a missing project file, are marked non-retriable.
• If any step fails, a compensating step releases the node and marks the job as failed, so a crashed render never keeps a machine reserved.
• The backend controls Windows hosts over async SSH: it creates a per-job folder, writes the nexrender manifest, starts the CLI and parses its output line by line to show live progress. Every remote call has a timeout.
• Heavy media never passes through the API: nodes download inputs and upload the MP4 straight to S3-compatible storage through short-lived presigned URLs.
• Slow AI services are polled on a fixed interval with a hard timeout, and B-roll clips are generated in parallel, at most four per video, to keep costs under control.

Stack

• Python
• FastAPI
• PostgreSQL
• SQLAlchemy
• Inngest
• asyncssh
• Adobe After Effects
• ExtendScript
• nexrender
• FFmpeg
• S3-compatible storage
• React
• TypeScript
• Docker

What this shows

• A desktop tool, After Effects, turned into a headless API service that produces videos without manual design work.
• Designers change templates without backend changes, because the layer contract and in-app scripts separate creative work from code.
• A cloud control plane reliably drives a remote Windows render farm with retries, compensating actions and progress reporting.
• Several AI services (LLM, avatar and text-to-video) combined into one pipeline with cost limits.

Questions about this case