You shouldn't have to choose between simple and powerful. Or between the cloud and control.
So we stopped choosing. Sandcastle is a source-available orchestrator for AI agents — define a workflow in one file, run it on any model, ship it anywhere on your own rules, and let it get sharper every run.
$pip install sandcastle-aiBuild
anything.
Complex agents from one simple file. 25 step types — parallel fan-out, conditions, loops, gates, deterministic $0 steps. Or skip the YAML entirely and describe what you want in a sentence; Sandcastle designs the workflow.
Any
model.
Model-independent by design. Claude, GPT, Mistral, Gemini, Llama — or fully local. Swap providers in one line. And when one goes down, auto-failover moves to the next while you sleep.
Anywhere,
your rules.
Cloud, your own server, fully isolated or air-gapped, or EU-only. The same workflow, your gate. EU data residency is built in — when you enable it, every model call stays in region.
- Cloudready
- Your server (self-host)ready
- Isolated / air-gappedready
- EU-only residencyenabled
It
learns.
It self-tunes on what succeeds and heals what fails. Every run is recorded — so you can replay your past workflows against new models and keep the winner. Smarter over time, on autopilot.
No feature gates. One version. Every line above ships free.
Sandcastle is a source-available orchestrator for AI agent workflows. A workflow is defined in a single YAML file and runs as a DAG with 25 step types — parallel fan-out, conditions, loops, gates, and deterministic steps that cost nothing to run. It connects to 7 model providers, cloud or local, and to 64 tool integrations. Every run is recorded and can be replayed from any step, a cost budget stops a run before it overspends, and the whole thing self-hosts on your own infrastructure with EU data residency. The source is available under BSL 1.1 and converts to Apache 2.0 in 2030.
The declaration ends. The build begins.