GitHub - dagploy/dax: AIOps Infra for deploy and manage self-hosted local AI in your own cloud. Vibe coding and AI agents compatible.

DAX: AIOps Infra as Code

Build and operate AI infrastructure inside your own cloud with YAML-based workflows at scale. Automate inference, training, and AI agent harnesses in real production environments. Supports spot instances, GPU quota-aware region switching, vibe-coding customization, and more.

Supported Cloud Providers

Google Cloud Platform (✅)
AWS (future development))
Azure (future development)

CLOUD PROVIDER: GCP

Pre-requisites: Enable GPU quota in your cloud project as early as possible. Approval can take up to 48 hours. Without GPU quota, launching GPU VMs may fail with a GPUS_ALL_REGIONS quota error. To reduce capacity issues, enable GPU quota across multiple regions.

⚡ 5 Minutes Setup

This step installs DAX on the default network without a public IP. Cloud NAT is required to enable internet access from inside the VM. You can log in to the VM with gcloud compute ssh <instance_name>.

1. Create a Service Account (~30 secs)

A service account is required as the owner/executor for provisioning instances, firewalls, and other services. Run this script to set it up. Make sure gcloud is installed and authenticated in your terminal.

bash <(curl -fsSL https://raw.githubusercontent.com/dagploy/dax/refs/heads/main/scripts/gcp_create_service_account.sh)

You will see the new service account created with required permission:

"roles/compute.instanceAdmin.v1"
"roles/compute.securityAdmin"
"roles/iam.serviceAccountUser"
"roles/artifactregistry.writer"
"roles/storage.objectUser"
"roles/compute.loadBalancerAdmin"
"roles/dns.admin"
"roles/secretmanager.secretAccessor"

This will produce both local service account JSON and secret dax-service-account-key that will use for provisioning any VM compute.

2. Setup Cloud NAT (~30 secs)

DAX server VM will have no public IP. To enable internet access for downloading packages, we create a cloud NAT

bash <(curl -fsSL https://raw.githubusercontent.com/dagploy/dax/refs/heads/main/scripts/gcp_install_cloud_nat.sh)

3. Create DAX VM service (~30 secs)

Run the command below. Replace YOUR-SERVICE-ACCOUNT-EMAIL with the service account email address you created earlier. You can find it in the generated service account JSON file.

Use --metadata enable-oslogin=TRUE to restrict access to OS Login, such as a corporate Google account. Use enable-oslogin=FALSE for standard SSH-based access.

gcloud compute instances create dax \
  --service-account=YOUR-SERVICE-ACCOUNT-EMAIL \
  --scopes=cloud-platform \
  --zone=us-central1-a \
  --machine-type=e2-custom-4-8192 \
  --boot-disk-size=60GB \
  --boot-disk-type=pd-balanced \
  --image-family=debian-12 \
  --image-project=debian-cloud \
  --network=default \
  --subnet=default \
  --no-address \
  --tags=dax \
  --metadata enable-oslogin=FALSE,startup-script='#!/bin/bash
set -e
apt-get update
DEBIAN_FRONTEND=noninteractive apt-get install -y git
'

4. Install DAX (~3 minutes)

SSH into the machine with gcloud compute ssh dax and run the installation step. DAX will be installed in your user folder.

sudo bash -c "$(curl -fsSL https://raw.githubusercontent.com/dagploy/dax/refs/heads/main/scripts/gcp_install.sh)"

Congrats, now DAX already installed and running 🎉

You can check the service with

sudo -iu dax -- tmux attach -t dax

💻 Connect with CLI

Any provisioning can be instructed to DAX server via curl or CLI. Connect your laptop/computer with DAX server via SSH tunnelling.

1. Install CLI

The detailed steps can be read here: Install DAX CLI (examples/project/dax-cli)

2. Tunnelling to DAX server

Run this command to establish connection securely over public internet. There are two ports: 8001 (DAX) and 8080 (Dashboard via Hatchet)

gcloud compute ssh dax --zone us-central1-a --tunnel-through-iap -- -L 8001:localhost:8001 -L 8080:localhost:8080

You can access the dashboard via https://localhost:8080 or curl provisioning into https://localhost:8001

EXAMPLE USE CASE

Run GPT OSS 20B in your cloud from scratch just takes 15 minutes.

Start by caching Docker images and models first — around 100GB in total — then launch the workload from the cache.

This cache mechanism can reduce startup time by up to 80% and lower costs by avoiding idle GPU time while large files are downloaded over the network.

Step 1: Cache the VLLM docker

dax run download_docker vllm/vllm-openai:nightly,ghcr.io/open-webui/open-webui:main --images vllm-lib --image-size 100

Step 2: Cache GPTOSS 20B from Huggingface

dax run download_hf openai/gpt-oss-20b --image-size 50

Step 3: Run the inference

dax run create_vm_inference --stack-name gptoss --config-json '{"images":["models--openai--gpt-oss-20b","vllm-lib"]}' --model openai/gpt-oss-20b

Or longer version

dax run create_vm_inference --stack-name gptoss --config-json '{"images":["models--openai--gpt-oss-20b","vllm-lib"]}' --model https://huggingface.co/openai/gpt-oss-20b

Access it from your laptop/computer via tunneling

gcloud compute ssh gptoss -- -L 8000:localhost:8000 -L 8081:localhost:8080

This will forwarding openwebui via http://localhost:8081 and VLLM API via http://localhost:8000

FAQ

1. My project is not changed

Property [project] is overridden by environment setting [CLOUDSDK_CORE_PROJECT. This is not DAX problem, but your local machine.

The solution: unset CLOUDSDK_CORE_PROJECT

2. Error launching: stack_name project_name program work_dir opts

local_workspace.py", line 1011, in create_or_select_stack
    raise ValueError(f"unexpected args: {' '.join(args)}")
ValueError: unexpected args: stack_name project_name program work_dir opts

Make sure the project path value defined in pulumi_yaml/Pulumi.yaml is correct.
Check if anything in .env is already correct.
Check on config/env/dev.yaml and make sure the value of project and service account is correct.

project_name: GCP_PROJECT_NAME
gcp:project: GCP_PROJECT_NAME
gcp:serviceAccount: SERVICE_ACCOUNT_EMAIL_ADDRESS

3. Error network

If you have problem with access to internet:

W: Failed to fetch https://deb.debian.org/debian/dists/bullseye/InRelease Cannot initiate the connection to`
debian.map.fastly.net:443 (2a04:4e42::644). - connect (101: Network is unreachable) Cannot initiate the connection to 
debian.map.fastly.net:443 (2a04:4e42:200::644). - connect (101: Network is unreachable) Cannot initiate the connection to

Or COS NVIDIA Driver installation stuck

Unable to find image 'us.gcr.io/cos-cloud/cos-gpu-installer:v2.7.2' locally
docker: Error response from daemon: Get "https://us.gcr.io/v2/": net/http: request canceled while waiting for connection (Client.Timeout exceeded while awaiting headers).
See 'docker run --help'.
Error: Failed to install GPU driver: could not install GPU drivers: failed to complete installation using installer 'us.gcr.io/cos-cloud/cos-gpu-installer:v2.7.2': exit status 125

This means the cloud NAT not working, the proxy haven't setup in correct way or the subnet haven't granted with google private access permission. Cloud NAT works at regional level, not global. To enable NAT and subnet, run this

bash scripts/gcp_install_cloud.nat.sh

DAX Cloud Services

We are working on the cloud services and AI Infra agents. If you are interested, you can join the waiting list or contact us for custom inquiry : https://www.dagploy.com/contact

Contributing

Visit CONTRIBUTING.md for information on building DAX from source or contributing improvements.

License

DAX is released under the Apache License 2.0. See LICENSE for the full text.

Citation

If you use DAX in your research, please cite:

@misc{dax,
  title = {DAX: AIOps Infra as Code},
  author = {DAGPLOY},
  year = {2026},
  url = {https://github.com/dagploy/dax}
}

推荐订阅源

Hacker News: Show HN