~17s
Default 8× GPU cloud burst
Splits your film at scene cuts across 8 GPUs in parallel and stitches closed H.264 GOPs with zero re-encoding.
Live benchmark29,676 frames rendering behind this page
Overcrank is on-demand cloud GPU rendering for videos built in code and Blender. Our engine is 33× faster on a single GPU — and because the cloud fans out across parallel GPUs for virtually the same cost, a 16-minute film that takes headless Chrome, Remotion, or HyperFrames 36 minutes comes back in under 20 seconds.
Full timeline163 × 183
Browser-based video tools step through your animation one screenshot at a time. Because a scripted video's timeline is already known up front, Overcrank computes motion across the entire film in advance and composites frames in parallel on the GPU.
Direct to NVENC1080p & 4K
Static HTML/CSS layers rasterize once, 3D scenes render headlessly over Vulkan, and finished frames go straight from GPU memory into hardware video encoders with audio mixed in.
Automatic verificationPixel-checked
Overcrank automatically spot-checks rendered scenes against headless Chrome. If any scene uses a browser feature outside the GPU fast path, Chrome renders just that scene so your final MP4 never breaks.
Benchmarks16.5-minute 1080p film · 103 scenes · 29,676 frames
~17s
Splits your film at scene cuts across 8 GPUs in parallel and stitches closed H.264 GOPs with zero re-encoding.
125×
Already 33× faster on a single machine (~65 s on one L40S or local RTX GPU) before multi-GPU fan-out even starts.
~6–9¢
8 GPUs running for 17 seconds use nearly the exact same GPU-seconds as 1 GPU running for 65 seconds.
9.5s
Preview a 12-second 4K Blender shot faster than playback on a warm GPU pool, or fan out full Cycles across 10 GPUs.
Benchmark details: Single-GPU Overcrank runs take 62–81 s (~65 s median) inside 1× warm NVIDIA L40S; 8× GPU parallel fan-out (~17 s active GPU wall time) partitions scenes across 8 L40S workers with parallel audio mix and lossless IDR stream-copy muxing (cold GPU spin-up adds a few seconds unless pre-warmed with overcrank warm). Browser baseline is hyperframes render (headless Chrome) on 16 CPU cores (2,146 s). Visual fidelity across 392 sampled frames: 38.39 dB median PSNR, zero frames flagged.
Built for AI agentsClaude Code · Cursor · Codex · Windsurf
Set up Overcrank so you can render our video and Blender projects on cloud GPUs:
1. Install the CLI + MCP server: `pip install "overcrank[mcp]"`
2. Run `overcrank llms` and save its rules as a skill so you know how to estimate costs, run fast previews, and verify renders.
3. Register `{"mcpServers": {"overcrank": {"command": "overcrank", "args": ["mcp"]}}}` in your MCP config.
4. Ask me for my Overcrank endpoint and API key (`overcrank config set endpoint https://api.overcrank.cloud` and `overcrank config set token`).
Whenever we edit an HTML/CSS/GSAP scene or a `.blend` file:
- Call `warm_up()` when you start editing so a GPU is hot before we export.
- Quote cost first, render a fast 5-second or preview-quality clip so we can inspect frames immediately, then render the full verified MP4 with `--max-cost`.
Works with any MCP-compatible agent or plain shell access. No local NVIDIA GPU, CUDA toolkit, or Blender install needed on your laptop.
Agents write GSAP timelines or tweak Blender lighting, render a preview in ~10 seconds, inspect dumped PNG frames, and fix layout or lighting issues autonomously.
Every render supports --estimate before uploading a single byte and --max-cost to halt automatically if a job ever exceeds your budget.
Change one caption or camera keyframe and only the modified file uploads. Unchanged assets and Chrome verification results stay cached.
# 1. quote time & cost before uploading
$ overcrank render ./my-video --estimate
# 2. fast 5-second draft while iterating
$ overcrank render ./my-video --seconds 5 --fallback never -o look.mp4
# 3. full Chrome-verified export, capped at $0.50
$ overcrank render ./my-video --max-cost 0.50 -o final.mp4
Interactive estimatorgrounded in logged GPU runs
Default 8× GPU fan-out splits scenes in parallel and stitches closed GOPs with zero re-encoding. $0.00 when idle.
Supported Pipelinesweb · 3d · blender
Web Video
Built for programmatic motion graphics, HyperFrames projects, and Remotion-style agent workflows. Landscape 16:9, vertical 9:16 Shorts/Reels, or custom dimensions with multi-track audio and video layers.
Real-Time & Ray-Traced 3D
Drop Three.js WebGPU scenes right between crisp HTML typography layers on a shared timeline, or switch to hardware ray-traced global illumination (ThreeRT) when shots call for real bounce light and soft shadows.
3D Film
.blend) — with live warm previewsPoint Overcrank at a .blend file and get an MP4 straight from Cycles on datacenter GPUs. Warm background Blender pools reload scene edits in place so 4K lighting and camera previews come back faster than the clip plays. At full sample counts, fan out shots across multiple GPUs.
| Quality preset | 1× L40S | 10× L40S |
|---|---|---|
| preview (sub-realtime) | 38 s | — |
| fast (36.4 dB) | 3 min 15 s | 31 s |
| full (37.8 dB) | 8 min 39 s | 61 s |
Architecture33× per GPU · 125× fanned out
A web browser recalculates style, layout, and paint 30 to 60 times every second. Overcrank evaluates your entire animation timeline once up front into compact per-frame transform matrices.
Each static text block, card, SVG, or image layer is laid out and rasterized once by a native Rust HTML/CSS engine instead of being redrawn thousands of times.
GPU kernels composite frame batches straight into hardware NVENC encoders — flying on a single local RTX card, or fanning out at scene boundaries across 8 cloud GPUs with lossless IDR stream stitching for the same compute cost.
Key frames from every scene are automatically compared against headless Chrome. Any scene using unsupported CSS or custom canvas scripts seamlessly falls back to Chrome.
Private Previewcurrent coverage
Since you opened this page00:00
Headless Chrome is on frame 0 of 29,676.