basilica-serverless-ops

name	basilica-serverless-ops
description	Use when the user wants to deploy apps, APIs, inference endpoints, or OpenClaw/Tau-style services with the Basilica CLI.

Basilica Serverless Ops

Use this skill for the managed deployment surface exposed by the CLI:

• generic basilica deploy
• inference templates like vllm and sglang
• basilica summon openclaw
• basilica summon tau
• deployment logs, status, scale, and delete
• share-token operations for private deployments

Choose The Right Deployment Mode

Use serverless deploys when the user wants:

• a public URL
• an HTTP API
• an inference endpoint
• an app that should come up without managing a box manually

Prefer rentals instead when the user needs SSH, custom system control, distributed training, or very large models that may spend a long time loading.

Canonical Preflight

Authenticate:

basilica login

Check spendability:

basilica balance

Before you create a deployment, decide:

• source form: inline code, Python file, or container image
• whether it should be public or --private
• whether it needs GPU
• whether it needs storage
• whether it should auto-expire via --ttl

Generic Deploy Patterns

Inline Python:

basilica deploy 'print("hello")' --name hello --port 8000 --ttl 300

Python file:

basilica deploy my_api.py \
  --name my-api \
  --port 8000 \
  --pip fastapi uvicorn \
  --ttl 600

Container image:

basilica deploy nginxinc/nginx-unprivileged:alpine \
  --name nginx-demo \
  --port 8080 \
  --cpu 250m \
  --memory 256Mi \
  --ttl 300

GPU-backed deploy:

basilica deploy inference.py \
  --name gpu-model \
  --gpu 1 \
  --gpu-model H100 \
  --memory 32Gi \
  --pip torch

Persistent storage:

basilica deploy hello.py \
  --name stateful-app \
  --storage \
  --storage-path /data

Deployment Management

List:

basilica deploy ls
basilica deploy ls --json

Status:

basilica deploy status my-app
basilica deploy status my-app --show-phases

Logs:

basilica deploy logs my-app
basilica deploy logs my-app --follow
basilica deploy logs my-app --tail 100

Scale:

basilica deploy scale my-app --replicas 3

Delete:

basilica deploy delete my-app --yes

Always use deploy delete, not stale deployments delete variants from older examples.

Inference Templates

vLLM:

basilica deploy vllm Qwen/Qwen2.5-0.5B-Instruct --name my-llm

SGLang:

basilica deploy sglang Qwen/Qwen2.5-0.5B-Instruct --name my-sglang

For large or unusual models, expect to tune health checks, GPU count, memory, or move to rentals.

OpenClaw And Tau

OpenClaw:

basilica summon openclaw --provider openai
basilica summon openclaw --provider anthropic

Tau:

basilica summon tau

OpenClaw-specific flags exposed by the current template include:

• --provider openai|anthropic
• --backend-url
• --model-id
• --provider-id
• --provider-api
• --context-window
• --max-tokens

OpenClaw template env requirements:

• OPENAI_API_KEY for OpenAI-backed deploys
• ANTHROPIC_API_KEY for Anthropic-backed deploys

Tau template env requirements:

• TAU_BOT_TOKEN
• CHUTES_API_TOKEN

Optional Tau env:

• TAU_CHAT_MODEL

Private Deployments And Share Tokens

Deploy privately:

basilica deploy my_api.py --name private-app --port 8000 --private

Manage share tokens:

basilica deploy share-token status private-app
basilica deploy share-token regenerate private-app
basilica deploy share-token revoke private-app --yes

Current behavior:

• deployments are public by default
• --private makes the deploy share-token-gated
• OpenClaw is a special case and is intended to be public with its own gateway token behavior

Topology And Placement Controls

Spread replicas across nodes:

basilica deploy hashicorp/http-echo:latest \
  --name spread-demo \
  --port 5678 \
  --replicas 2 \
  --unique-nodes

Use explicit spread policy:

basilica deploy image \
  --name app \
  --replicas 4 \
  --spread-mode required \
  --max-skew 1

Operational Caveats

• Use --ttl by default unless the user explicitly wants a persistent service.
• Large models can spend 10-30 minutes loading. Watch deploy status --show-phases and deploy logs --follow.
• If a model is huge enough to need 8x H100/H200 or lots of manual tuning, prefer rentals first.
• Deploys are public unless you say otherwise.

Good Agent Flow

1. check basilica balance
2. pick the deploy shape
3. deploy with --ttl unless persistence is required
4. watch status/logs until the URL is usable
5. return the URL and the cleanup plan
6. delete the deployment when done unless the user asked to keep it

File Pointers

• examples/15_cli_deploy/README.md
• examples/inference/README.md
• crates/basilica-cli/src/cli/handlers/deploy/mod.rs
• crates/basilica-cli/src/cli/handlers/deploy/templates/openclaw.rs
• crates/basilica-cli/src/cli/handlers/deploy/templates/vllm.rs
• crates/basilica-cli/src/cli/handlers/deploy/templates/sglang.rs
• crates/basilica-cli/src/cli/handlers/deploy/templates/tau.rs

TODOs

• add a repo-local end-to-end example for summon openclaw with provider env setup
• add a tested table mapping model size to recommended deploy vs rentals path