Agent-first compute for AI agents and machine learning workloads.
Your agent provisions compute, runs the job, and returns the result.
Early access. Programmable through CLI and REST API.From training to evaluation. Follow a worker from setup to saved report.
A clean Ubuntu starting point. Bring the tools and code your task needs.
Add your tools. Run your code.A clean Ubuntu starting point. Bring the tools and code your task needs.
zeist box create eval-worker --waitThe sample shell command creates workspace after the runtime is running. Your application and dependencies are added after creation.
Take the evaluation worker from a fresh environment to an independent experiment.
zeist box create eval-worker --waitAllocate compute and a root disk, ready for your code.
This illustrative sandbox is named eval-worker. The command waits for the environment to be ready.
The worker can stop. Its report doesn’t disappear. Keep results in a Volume or connect your own bucket.
The runtime ends. report.json stays in run-results for the next run.
eval-worker writes through a writable /mnt/results mount; eval-next reads the same Volume at /mnt/data. Managed Volumes are persistent object storage and survive sandbox deletion. Use a writable prefix for each run, such as training-a/. Publish writes with fsync before sharing or forking, and use one writer per file or prefix.
report.jsonSaved reporteval.jsonlEvaluation inputGive the same runtime a different job—from evaluating records to training a policy.
eval-workerzeist box exec eval-worker -- python agent.pyfile.read/mnt/data/eval.jsonlshell.executepython evaluate.pyfile.write/mnt/results/report.json/mnt/resultsIllustrative run · sample dataYour agent reads eval.jsonl, runs evaluate.py, and saves report.json to run-results.
The example assumes your script and its dependencies are available in eval-worker. Read mounted data at /mnt/data and write artifacts to a writable mount at /mnt/results.
The agent uses file.read, shell.execute, and file.write as illustrative tools for reading records, running the evaluation, and saving its report.
Package the evaluation worker in a private VM template. Keep a version. Launch the same setup for the next run.
WORKDIR /workspace
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY evaluate.py .requirements.txtDependenciesevaluate.pyYour codeApplication-layer example. Template builds use the supported Zeist Ubuntu base.
eval-worker:v1eval-runReadyevaluate.pyeval-worker:v1The template stays. Each run starts fresh.
zeist box create eval-run --template eval-worker:v1 --waitLaunch eval-run from eval-worker:v1. The private template stays available.
Set outbound domain and CIDR rules. Private CLI tunnels expose services through your local machine; public HTTPS endpoints allow unauthenticated access.
A few things to know about Zeist today.
Zeist is in early access. Join the waitlist to request access and tell us what you are building.
Run AI agents, reinforcement learning experiments, model evaluations, and data preparation workloads. Start from Ubuntu Base, Linux Desktop, a public Linux container image, a private template, or a saved checkpoint. Bring the tools and dependencies your work needs.
Managed Volumes preserve data beyond sandbox deletion. Pause and checkpoints save the root disk. A fork copies the running sandbox, while external Volume and Bucket mounts remain shared. Mounts are read only by default.
Yes. Connect an existing bucket from AWS S3, Google Cloud Storage, Hetzner, Tigris, or Cloudflare R2 and mount it into your runtime. Your bucket stays with its provider. Removing the Zeist connection does not delete the upstream bucket or its data.
The current timeout limit is 24 hours. Extend a sandbox within that limit, or restore a saved checkpoint into a new sandbox for the next session.
Public HTTPS endpoints are unauthenticated. For access through your local machine, use a private CLI tunnel instead. Outbound domain and CIDR rules let you control network access.
Yes. Use the CLI or REST API to create environments, execute code, and manage runtime resources programmatically. The dashboard lets you inspect and manage those same resources. Zeist is in early access.