01 / 06
“Every run needs someone who knows SLURM.”
Build it visually
Drag stages onto a canvas. No DSL, no SLURM scripts. Anyone on your team can launch it.
Horus for data pipelines, simulation and AI training
Your data, accelerated.
Run your data and compute pipelines across HPC clusters, cloud GPUs and your own machines. Your team gets results without waiting on engineering.
Backed by and running at
The problem
The bottleneck isn't your team. It's the setup around it.
“For my team to use HPC I need DevOps engineers, and I can't find them.”
Head of R&D “To launch a run I depend on someone else having time.”
Simulation engineer “Installing anything on the cluster means a security review.”
HPC admin “Only the person who wrote the pipeline understands it.”
Data science lead “Every loop from idea to result takes too long.”
R&D director
Horus is the engineering layer you don't have.
It writes the jobs, moves the files and records every result. Your team does the part only it can do.
How Horus solves it
Build it, launch it, trace it. Horus does the plumbing.
02 / 06
“The notebook works. Getting it onto the cluster does not.”
Bring your notebook
Decorate a Python function as a task. The same code runs on HPC or cloud.
pipeline.py
from horus_builtin.task.function import FunctionTask
@FunctionTask.task(wf, target=cluster)
def simulate(case):
return solve(case)
simulate · running on cluster
03 / 06
“Half my afternoon is scp.”
Files move themselves
Inputs and outputs are staged over SSH between your machine, the cluster and the cloud.
Your machinessh
HPC clusterssh
Cloudssh
cases.csv→ cluster
results.parquet→ cloud
model.pt← your machine
04 / 06
“We pay GPU rates to prep data.”
Each step on the hardware it needs
CPU here, GPU there. HPC or cloud, chosen per stage.
Local
anyHPC
CPUCloud
GPU
anyHPC
CPUCloud
GPU
Prepare inputs
CPU
Simulate
CPU
Train model
GPU
Report
CPU
05 / 06
“Ten thousand jobs means ten thousand terminals.”
Sweeps at scale
Launch 10,000 simulations and track them from one place.
Parameter sweep10,000 jobs
donerunningqueued
06 / 06
“Nobody remembers which parameters made that plot.”
Every result traceable
Lineage from a plot back to the input data, parameters and machine. Cached artifacts are reused, not recomputed.
sweep_plot.png
result
model.pt
train · cloud GPU
results.parquet
simulate · HPC
params.yaml + cases.csv
inputs
For IT and HPC
Your cluster stays exactly as it is.
Horus reaches it over plain SSH. No daemons, no agents, nothing new for IT to review.
Plain SSH
Horus talks to SLURM the way you do. No agent, no daemon, no service account.
Your allocation
Hours you already have get used first. Cloud only where you want it.
your laptop
$ horus run sweep.yaml
→ ssh login01.hpcconnected
→ stage params.yaml cases.csv2 files → /scratch/run-4f2a
→ sbatch simulate.shjob 81734 submitted
RUNNING simulatenode-c17
COMPLETED simulateexit 0
← pull results.parquet1 file
✓ clean /scratch/run-4f2atemp files removed
0 agents installed
on the cluster. Horus speaks plain SSH.
Your containers
The Docker and Apptainer images your team already builds.
Your data
You choose which stage leaves the cluster, and which never does.
Hybrid, by default
You already pay for the infrastructure. Let Horus use 100% of it.
One workflow fills your workstation, your GPU servers and your SLURM allocation to capacity. Only the work that does not fit spills over, to another machine of yours or to the cloud.
parameter-sweep0 queued
run log
- $ horus run sweep.yaml
- → map simulate × 24workstation
- Full, spilling...gpu-server
- Full, spilling...slurm
- Provisioning cloud node...cloud
- ✓ 24/24 placed0 idle
0 idle slots
on machines you already pay for, before the cloud starts.
Your machines, one pool
Workstations, GPU servers and your SLURM allocation take items from the same workflow.
Filled, never oversubscribed
Each machine runs as many items as its capacity allows. Extra items wait for a free slot.
Spill only what is left
When a machine is full, work spills to the next one you allowed: another of yours, or the cloud.
Cloud, on demand
Machines that exist only while your run does
When a stage needs a GPU, Horus creates the machine as the run reaches it and destroys it when the run ends. You pay in tokens only while it exists.
parameter-sweep· laptop + SLURM + cloud
run log
- $ horus run sweep.yaml
- → Prepare inputslaptop
- Provisioning cpu nodes...Simulate
- Node ready.Simulate
- Provisioning gpu node...Train model
- Node ready.Train model
- → Reportlaptop
- ✓ Run completed
- Destroying node...Simulate · Train model
- Node destroyed.Simulate · Train model
0 machines left running
after the run ends. Billing stops when they are destroyed.
No cloud account
Horus creates the machine. You pick the CPUs, GPU and memory.
Mix with your cluster
Simulate on SLURM, burst to the cloud, train on a cloud GPU.
Nothing left running
Torn down when the run ends: success, failure or cancel.
Who it's for
One platform. Three people who stop waiting.
Head of R&D
More experiments from the team you have.
“Every new idea needs an engineer first.”
- More iterations with the same team, no new hires.
- Traceability built in: trace any result back to its inputs, parameters and machine.
IT and HPC admins
Approve it once. Every team reuses it.
“Every new tool means a security review and a queue of tickets.”
- Plain SSH and your existing allocation.
- Workflows validated once, reused by every team.
Researchers and engineers
Launch it yourself.
“I wait on someone else, then move files by hand.”
- Build in the visual builder, or add @task to a Python function.
- Real-time status for every stage.
Life sciences and chemistry
Sensitive stages in house, heavy compute outside. Every result traced to its inputs.
Engineering and simulation
Thousands of solver runs across your cluster and the cloud, from one workflow.
AI and data teams
Prep data on CPUs, train on GPUs on demand. Nothing idle between runs.
Research computing centres
Self-service for every team, on the cluster you already run.
Agents can launch and track real workflows through MCP, with full traceability.
Examples
Real pipelines, ready to run
34 documented pipelines. Each one shows which stage runs where, and why there.
- Prepare inputslaptop · CPU
- Simulate ×240SLURM · HPC
- Train modelcloud · H100
- Reportlaptop · CPU
horus-runtime: free and open source
The engine under Horus. Install it today. AGPL v3.
Questions
Do I need to install anything on my cluster?
No. Horus connects over plain SSH and stages files itself. No Horus software is installed on the cluster.
Does my team need to know SLURM?
No. They build pipelines in the visual builder or with a Python @task decorator. Horus writes and submits the jobs.
Can one workflow use HPC and cloud GPUs?
Yes. Each stage declares its own resources, so every step runs on the hardware it needs: CPU, GPU, HPC or cloud.
Do I need my own cloud account?
No. For a stage you send to the cloud, such as a GPU stage, Horus creates the machine you specify when the run reaches it and destroys it when the run ends. You pay in Horus tokens only while it exists.
Is Horus open source?
The engine is. horus-runtime is on GitHub and PyPI under AGPL v3.
Can AI agents run workflows?
Yes. Horus exposes workflows over MCP, so agents can build, run and inspect them.
Run completed · 4/4
Your next result should not wait on infrastructure.
Open Horus and run your first workflow, or start with the open-source runtime today.