What sandboxes are for

Some Qyra features use an AI agent (Claude Code) that writes and runs code on your behalf. Today that's:

• AI writeback — the agent edits your dbt project (e.g. adds a metric or dimension), runs qyra compile to validate it, and opens a pull request.
• Data app generation — the agent generates and builds a small web app from a prompt.

Running model-generated code directly on the Qyra server would be unsafe: the code is untrusted, can run arbitrary commands, and needs its own toolchain (git, dbt, the Qyra CLI, Node). Qyra instead runs each agent inside a sandbox — an isolated, disposable environment with a constrained network. The agent does its work there, Qyra collects the result (a PR, a built app), and the sandbox is torn down.

Sandboxes are also what make these features fast and multi-turn: a sandbox can be suspended between turns and resumed later, so a conversation with the agent keeps its state without holding a container open the whole time.

Sandbox providers

The sandbox backend is pluggable. Qyra talks to a provider-neutral interface, so the same feature code runs on whichever backend your deployment is configured for. You select the provider with the SANDBOX_PROVIDER environment variable.

E2B, AWS Lambda MicroVMs, and Azure Container Apps Sandboxes are all supported production backends. E2B is the managed default; the AWS and Azure providers are for teams who want sandboxes to run inside their own cloud account. More providers (Kubernetes, ECS) are planned.

LLM gateway egress

Data apps can route Claude or Codex model traffic through a corporate LLM gateway. The sandbox connects to that gateway independently from the Qyra backend, so both network paths must be allowed.

• E2B and Azure Sandboxes receive the hostname from ANTHROPIC_BASE_URL, BEDROCK_BASE_URL, or OPENAI_BASE_URL in their per-sandbox egress allowlist.
• AWS Lambda MicroVMs cannot apply Qyra's per-sandbox allowlist. If you use a restricted LAMBDA_MICROVM_EGRESS_CONNECTOR_ARN, configure that connector to permit the gateway hostname.
• Docker and Google Cloud Run use their existing runtime network path. Make sure its DNS and firewall policy can reach the gateway.

These settings affect Data app coding agents, not AI writeback. If AI writeback is enabled, its sandbox still needs its existing Anthropic and git-host egress.

See Corporate LLM gateways for provider configuration and the required API paths.

E2B (production default)

E2B runs each sandbox as a Firecracker microVM in E2B's cloud. It's the default — if you don't set SANDBOX_PROVIDER, Qyra uses E2B.

To use it you need an E2B account and API key, and the agent needs an Anthropic API key:

SANDBOX_PROVIDER=e2b            # default, can be omitted
E2B_API_KEY=e2b_...            # from your E2B dashboard
ANTHROPIC_API_KEY=sk-ant-...   # the agent (Claude Code) runs inside the sandbox

The sandbox images are E2B templates. Qyra uses separate templates for data apps and for AI writeback so they can be pinned or rolled back independently. These default to the published Qyra templates and rarely need to be set:

E2B_TEMPLATE_NAME=qyra/qyra-data-app
E2B_TEMPLATE_TAG=<qyra-version>             # defaults to your Qyra version
E2B_AI_WRITEBACK_TEMPLATE_NAME=qyra/qyra-ai-writeback
E2B_AI_WRITEBACK_TEMPLATE_TAG=<qyra-version>

AWS Lambda MicroVMs (self-hosted production)

AWS Lambda MicroVMs run each sandbox as a Firecracker microVM inside your own AWS account, so untrusted agent code and your repository contents never leave your infrastructure. The microVMs have no public IP — your backend reaches each one through an AWS-managed endpoint that requires a short-lived per-microVM token — and you control their outbound network access (see Networking and IAM).

This is the recommended sandbox provider for customers deploying Qyra on AWS — it keeps the sandbox boundary inside your existing AWS account and avoids sending agent workloads or repository contents to a third-party service.

Prerequisites

Provision these with your own IaC, in the same AWS account and region your Qyra backend already runs in:

• Two MicroVM images — one for data app generation and one for AI writeback (they bundle different toolchains). Build them from the Dockerfiles in the Qyra repo (sandboxes/data-apps/, sandboxes/ai-writeback/, and the exec agent in sandboxes/microvm-agent/), push them to ECR, and register each as a Lambda MicroVM image on the AWS-managed al2023 base — ARM_64, 4 GB memory, with the agent's /ready hook on port 8080. Each registration returns an image ARN for the config below.
• Control-plane permissions on your backend's existing IAM role — add RunMicrovm, GetMicrovm, SuspendMicrovm, ResumeMicrovm, TerminateMicrovm, and CreateMicrovmAuthToken.

Configure the provider

SANDBOX_PROVIDER=lambda-microvm
ANTHROPIC_API_KEY=sk-ant-...   # the agent (Claude Code) runs inside the microVM

# Region the microVMs run in (defaults to eu-west-1, the EU launch region)
LAMBDA_MICROVM_REGION=eu-west-1

# The image ARNs from registering the two MicroVM images above. Required.
LAMBDA_MICROVM_DATA_APP_IMAGE_ARN=arn:aws:lambda:<region>:<account>:microvm-image/...
LAMBDA_MICROVM_AI_WRITEBACK_IMAGE_ARN=arn:aws:lambda:<region>:<account>:microvm-image/...

The backend uses its ambient AWS credentials (instance role / IRSA / standard SDK credential chain) to call the Lambda MicroVMs control plane, so no access keys are configured here.

Networking and IAM

Override these to tighten the network boundary or to give the microVM an IAM role:

# IAM role the microVM assumes (e.g. to pull from a private ECR or reach AWS APIs)
LAMBDA_MICROVM_EXECUTION_ROLE_ARN=arn:aws:iam::...:role/qyra-sandbox

# Network connectors. Default to AWS-managed open ingress/egress; point these at
# your own VPC connectors to constrain traffic. We strongly recommend an egress
# connector that denies outbound traffic by default.
LAMBDA_MICROVM_INGRESS_CONNECTOR_ARN=arn:aws:lambda:<region>:aws:network-connector:aws-network-connector:ALL_INGRESS
LAMBDA_MICROVM_EGRESS_CONNECTOR_ARN=arn:aws:lambda:<region>:aws:network-connector:aws-network-connector:INTERNET_EGRESS

Azure Container Apps Sandboxes (self-hosted production)

Azure Container Apps Sandboxes run each sandbox as an isolated, microVM-class environment inside your own Azure subscription, so untrusted agent code and your repository contents never leave your infrastructure. Sandboxes have native suspend/resume (a full memory + disk snapshot with sub-second restore), which is what keeps multi-turn agent conversations fast.

This is the recommended sandbox provider for customers deploying Qyra on Azure — it keeps the sandbox boundary inside your existing Azure tenant and avoids sending agent workloads or repository contents to a third-party service.

Prerequisites

Provision these in the same Azure subscription and region your Qyra backend runs in, using the aca CLI or the Sandboxes portal:

• A sandbox group per feature — one for data app generation and one for AI writeback (they bundle different toolchains). A sandbox group (Microsoft.App/SandboxGroups) is the management boundary that holds a feature's sandboxes and disk image. Give each group a Memory-mode auto-suspend lifecycle policy so idle sandboxes snapshot and scale to zero.
• A disk image per group — build the two images from the Dockerfiles in the Qyra repo (sandboxes/data-apps/, sandboxes/ai-writeback/), push them to a container registry (e.g. Azure Container Registry), and register each as a disk image in its sandbox group. Registration returns a disk image ID for the config below. (Unlike the AWS provider, there is no in-VM agent to build — Sandboxes expose a native command/file API.)
• A workload identity with the data-plane role — grant your backend's managed identity the Container Apps SandboxGroup Data Owner role on each sandbox group. Qyra uses DefaultAzureCredential (workload identity on AKS, or the standard Azure credential chain) to authenticate — no client secret is configured here.

Configure the provider

SANDBOX_PROVIDER=azure-sandboxes
ANTHROPIC_API_KEY=sk-ant-...   # the agent (Claude Code) runs inside the sandbox

# Where your sandbox groups live
AZURE_SANDBOXES_SUBSCRIPTION_ID=<subscription-id>
AZURE_SANDBOXES_RESOURCE_GROUP=<resource-group>
AZURE_SANDBOXES_REGION=eastus2

# Per-feature sandbox group + disk image ID (from the prerequisites above). Required.
AZURE_SANDBOXES_DATA_APP_GROUP=qyra-data-app
AZURE_SANDBOXES_DATA_APP_DISK_IMAGE=<data-app-disk-image-id>
AZURE_SANDBOXES_AI_WRITEBACK_GROUP=qyra-writeback
AZURE_SANDBOXES_AI_WRITEBACK_DISK_IMAGE=<writeback-disk-image-id>

# Optional — sandbox size (XS/S/M/L, default M)
AZURE_SANDBOXES_RESOURCE_TIER=M

Egress is locked down automatically: each sandbox launches with a default-deny egress policy that only allows the hosts the agent needs (the Anthropic API and your git host), with full traffic inspection so the platform enforces the deny on all traffic and blocks non-HTTP egress. Untrusted agent code can't reach any other destination — outbound requests to non-allowlisted hosts are rejected, and all other ports are blocked.

Google Cloud Run Sandboxes (self-hosted, preview)

Use the gcp-cloud-run provider to run data app sandboxes in a Cloud Run service. Qyra connects to the service through an HTTP gateway included in the sandbox image.

Prerequisites

• A Google Cloud project with billing enabled.
• A recent version of the gcloud CLI that supports gcloud beta run deploy --sandbox-launcher.
• A checkout of the Qyra repository.
• External object storage configured for Qyra.
• An Anthropic API key.

1. Enable the Google Cloud APIs

gcloud services enable run.googleapis.com cloudbuild.googleapis.com \
  --project <your-project>

2. Build the gateway image

From the root of the Qyra repository, run:

cd sandboxes/data-apps

STAGE=$(mktemp -d)
cp -R template qyra-query-sdk.tgz "$STAGE/"
npm pack @computesdk/cloud-run --pack-destination "$STAGE" >/dev/null
tar -xzf "$STAGE"/computesdk-cloud-run-*.tgz \
  -C "$STAGE" package/dist/gateway.mjs
mv "$STAGE/package/dist/gateway.mjs" "$STAGE/gateway.mjs"

sed 's|^RUN chown -R user:user /app|RUN useradd -m -s /bin/bash user 2>/dev/null \|\| true|' \
  e2b.Dockerfile > "$STAGE/Dockerfile"

cat >> "$STAGE/Dockerfile" <<'EOF'

COPY gateway.mjs /gateway/gateway.mjs
ENV NODE_ENV=production
CMD ["node", "/gateway/gateway.mjs"]
EOF

gcloud builds submit "$STAGE" \
  --tag gcr.io/<your-project>/qyra-sandbox-gateway:latest \
  --project <your-project>

3. Deploy the gateway

Generate a shared secret and deploy the image:

SANDBOX_SECRET=$(openssl rand -hex 32)

gcloud beta run deploy qyra-sandbox-gateway \
  --image gcr.io/<your-project>/qyra-sandbox-gateway:latest \
  --region <region> \
  --project <your-project> \
  --sandbox-launcher \
  --no-cpu-throttling \
  --session-affinity \
  --cpu 4 \
  --memory 16Gi \
  --concurrency 20 \
  --min-instances 1 \
  --max-instances 1 \
  --timeout 3600 \
  --allow-unauthenticated \
  --set-env-vars "SANDBOX_SECRET=$SANDBOX_SECRET"

Save SANDBOX_SECRET securely. You need the same value when configuring Qyra. Although the Cloud Run service allows unauthenticated requests, the gateway requires this secret for every sandbox operation.

Keep the service at one instance. Active sandboxes run inside that instance and are lost if Cloud Run replaces it during a generation.

4. Configure Qyra

Get the gateway URL:

gcloud run services describe qyra-sandbox-gateway \
  --region <region> \
  --project <your-project> \
  --format='value(status.url)'

Set these environment variables on both the Qyra backend and scheduler:

SANDBOX_PROVIDER=gcp-cloud-run
ANTHROPIC_API_KEY=sk-ant-...
GCP_CLOUD_RUN_SANDBOX_URL=<gateway-url>
GCP_CLOUD_RUN_SANDBOX_SECRET=<sandbox-secret>

Restart the backend and scheduler to apply the configuration.

5. Restrict sandbox egress (recommended)

Sandbox traffic exits through the gateway instance's network path, so it follows the service's VPC egress setting and can be filtered with standard VPC networking. None of this is provisioned by Qyra — you configure it yourself:

1. A subnet on the VPC network you'll filter, in the same region as the gateway service.

2. A Cloud Router and Cloud NAT on that network. Once the service routes all traffic through the VPC, outbound internet access only works via NAT — without it, sandboxes and the gateway itself lose all egress. Reserve a static IP for the NAT if you want a stable egress IP to allowlist on your side (for example, on your git host):

   gcloud compute routers create sandbox-egress-router \
     --network <network> --region <region> --project <your-project>
   
   gcloud compute routers nats create sandbox-egress-nat \
     --router sandbox-egress-router --region <region> --project <your-project> \
     --auto-allocate-nat-external-ips --nat-all-subnet-ip-ranges

3. Direct VPC egress on the gateway service. This deploys a new revision, so any active sandboxes are lost — do it while no generation is running:

   gcloud beta run services update qyra-sandbox-gateway \
     --region <region> --project <your-project> \
     --network <network> --subnet <subnet> --vpc-egress all-traffic

   All sandbox traffic now leaves through your VPC: its public egress IP is the NAT IP, and VPC firewall egress rules apply to it — a deny rule for a destination range blocks matching sandbox traffic while the rest keeps flowing.

4. Filter what sandboxes can reach. VPC firewall rules give you IP/CIDR-level control. For a domain-level allowlist, add Secure Web Proxy: create a deny-all-egress firewall rule that only allows traffic to the proxy, and allowlist the domains the sandboxes and the gateway need — at minimum api.anthropic.com, registry.npmjs.org, and your object storage endpoint. Secure Web Proxy is an explicit proxy: clients must be configured to use it (bake proxy settings into the gateway image — sandboxes share its root filesystem), and the deny-all firewall drops anything that bypasses it.

To update the sandbox toolchain, rebuild the gateway image and redeploy the Cloud Run service.

Local Docker provider (development)

For local development you can run sandboxes as plain Docker containers on your own machine — no E2B account required. This is the recommended way to work on or try the AI features locally.

It uses the same images E2B builds, but as plain local Docker images. Two separate images are used (different toolchains), mirroring the two E2B templates:

Prerequisites

• Docker running locally, with the daemon reachable from the Qyra backend.
• S3-compatible object storage configured (locally this is MinIO). Suspended-sandbox snapshots are tarred to object storage so a conversation survives the container being destroyed — see external object storage.
• An Anthropic API key (ANTHROPIC_API_KEY) for the agent.

Setup

1. Build the local sandbox images (each builds from sandboxes/<feature>/):

   ./sandboxes/data-apps/build-local-image.sh        # -> qyra-sandbox:local
   ./sandboxes/ai-writeback/build-local-image.sh     # -> qyra-ai-writeback:local

   These are large (the writeback image bundles dbt, the Qyra CLI and Claude Code) and only need rebuilding when the sandbox toolchain changes.

2. Point Qyra at the Docker provider:

   SANDBOX_PROVIDER=docker
   # optional — these are the defaults:
   SANDBOX_DOCKER_IMAGE=qyra-sandbox:local
   SANDBOX_AI_WRITEBACK_DOCKER_IMAGE=qyra-ai-writeback:local

3. Restart the backend and the scheduler so both pick up the new environment. Data app generation runs in the scheduler worker, so a stale SANDBOX_PROVIDER there will keep it on E2B. (With PM2, a plain restart reuses the cached env — delete and re-start the processes, or restart with --update-env, to actually reload the env file.)

Snapshot lifecycle

Every turn suspends its own sandbox, so in steady state nothing sits idle. Two timers configure the cloud-side idle policy as a backstop — when to auto-suspend a sandbox left running and when to auto-terminate a suspended one:

SANDBOX_IDLE_TIMEOUT_MS=1800000        # auto-suspend a running-but-idle sandbox (default 30 min)
SANDBOX_SNAPSHOT_RETENTION_MS=604800000 # auto-terminate a suspended microVM (default 7 days); also how long a thread stays resumable

SANDBOX_IDLE_TIMEOUT_MS feeds the auto-suspend policy on both the Lambda MicroVMs and Azure Sandboxes providers (for Azure it sets each sandbox's Memory-mode auto-suspend interval). SANDBOX_SNAPSHOT_RETENTION_MS is read only by Lambda MicroVMs — on Azure, suspended-sandbox retention is governed by the sandbox group's own auto-delete policy. E2B manages idle sandboxes itself, and the Docker and Cloud Run providers have no cloud-side idle handling — both persist suspend-time snapshots to object storage and destroy the sandbox, so nothing idles in the first place.

Environment variable reference