The build system for AI generation.
Atlas is Git + Make for AI image and video: content-addressed so identical work is never repeated, resumable so a dead GPU never costs you progress, and budget-gated so you always know the price before you spend it.
$ atl render main t2i@1 2df0… running done t2i@1 2df0… -> 81a4… 1.2s $0.03 i2v@1 9fef… running done i2v@1 9fef… -> 1b47… 38s $0.36 2 tasks executed, 0 cached, $0.39 spent $ atl render main # run it again — nothing changed 0 tasks executed (all cached) · $0.00 spent
Generation that behaves like engineering.
The three ways AI pipelines waste your time and money — solved at the kernel, not bolted on.
Never re-render the same work
Every recipe is content-addressed. Identical inputs hash identically, so Atlas serves the cached artifact instead of paying for it twice. Caching is structural, not optional.
Never lose work to a kill
Runs are journaled and resumable. A dead RunPod, a killed process, a crashed laptop — atl resume picks up exactly where it stopped. Completed work stays done.
Know the cost first
Paid work refuses to run without a budget envelope covering it. atl plan shows the price before a single API call. No surprise bills, ever.
Author intent. Atlas handles the rest.
Describe the work as recipes
Text-to-image, image-to-video, compositing — declared as a graph of content-addressed recipes, by hand or from the Python Director's DSL.
Set a budget, see the plan
Open an envelope, run atl plan. You get the exact task list and a P90 cost estimate before anything spends.
Render — cached, resumable, accountable
Atlas runs only what's missing, records every artifact's model, seed, cost and license in a provenance ledger, and resumes cleanly if interrupted.
Change one thing, re-render only what's dirty
Bump a character design and atl dirty lists exactly which shots fell behind — and what fixing them costs. The rest stay cached.
$ atl canon bump mara.visual --ref sheet_v2.png mara.visual: v1 -> v2 $ atl dirty 2 shots reference an outdated design: shot/01/01 mara closeup re-render ~$0.39 shot/02/02 mara on train re-render ~$0.39 4 other shots unaffected — still cached. estimated to fix: $0.78
Free to start. Fair as you grow.
The core CLI is free and open source, forever. Pay only when Atlas is doing real work for a real team.
- Full render engine & cache
- Resumable, budget-gated runs
- Replicate API bridge
- Python Director's DSL
- Community support
- Everything in Core
- Pro plugin pack & premium model bridges
- Priority email support
- Early access to new capabilities
- Commercial use
- Everything in Pro
- Self-host commercial license
- Up to 10 seats
- Private plugin registry
- SLA support
- Hosted, resumable renders
- Usage-based billing
- Team dashboards
- In development
Prefer to just support the project? Sponsor Atlas on GitHub.
Up and running in two minutes.
macOS / Linux
$ brew install REPLACE_WITH_YOUR_ORG/tap/atl # or grab a binary from Releases $ atl init && atl --help
Windows
> scoop install atl # or download atl.exe from Releases > atl init; atl --help
Full quickstart and the plugin protocol are in the README and protocol docs.
Questions, answered.
Is Atlas really free?
Yes. The Core CLI is Apache-2.0 licensed and free forever. You only pay Atlas for Pro features, a commercial self-host license, or (later) managed Cloud rendering. You always pay your model provider (e.g. Replicate) directly for generation.
Does my API token or artwork ever touch your servers?
No. Atlas runs entirely on your machine. Your provider tokens are read from environment variables and never written to disk, logs, or any Atlas service. Plugins run as sandboxed subprocesses with a filtered environment.
Which models and providers are supported?
v0.1 ships a Replicate bridge for text-to-image and image-to-video. The plugin protocol is open and versioned, so any provider can be added as a plugin — the Pro plan bundles additional premium bridges.
What's "Atlas Cloud"?
A future managed service that runs your renders on our infrastructure with usage-based billing — the same resumable, budget-gated engine, without you managing GPUs. Join the waitlist to be first in.
Can I use Atlas commercially?
Yes. Core is Apache-2.0 (commercial use permitted). Pro and Team add commercial support and a self-host license for organizations that need one.
Stop paying twice for the same frame.
Install the CLI, open a budget, and render your first shot in minutes.