modal-deploy

name	modal-deploy
description	Deploy serverless Python functions and AI workloads to Modal. Use when user needs GPU, AI inference, serverless Python, or scheduled AI tasks.
argument-hint	[script-path]
disable-model-invocation	true
allowed-tools	Bash, Read, Write, Edit, Glob, Grep

Modal Deployment (Serverless Python + GPU)

Goal

Deploy Python functions to Modal for serverless execution. $30/month free credit. GPU available for AI tasks.

When to Use

• AI/ML inference (LLM calls, embeddings, image generation)
• Serverless Python functions (run on demand, scale to zero)
• GPU workloads (T4, A100, H100)
• Scheduled AI tasks (daily RAG ingestion, batch processing)
• Webhook endpoints for AI processing

When NOT to Use

• Always-on bots (use Koyeb)
• Frontend/dashboard (use Vercel)
• Visual workflows (use n8n)

Prerequisites

• Modal account (modal.com — free, $30/mo credit)
• pip install modal
• modal setup (one-time auth)

Step-by-Step

1. Install and authenticate

pip install modal
modal setup
# Opens browser for auth — click approve

2. Create a Modal app

# app.py
import modal

app = modal.App("my-automation")

@app.function()
def generate_post(topic: str) -> str:
    """Generate a social media post using AI."""
    from openai import OpenAI
    import os

    client = OpenAI(
        base_url="https://api.euron.one/api/v1/euri",
        api_key=os.environ["EURI_API_KEY"],
    )
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role": "user", "content": f"Write a LinkedIn post about {topic}"}],
    )
    return response.choices[0].message.content

# Webhook endpoint (callable via HTTP)
@app.function()
@modal.web_endpoint(method="POST")
def webhook(data: dict):
    topic = data.get("topic", "AI automation")
    post = generate_post.remote(topic)
    return {"post": post}

# Scheduled task (runs daily)
@app.function(schedule=modal.Cron("0 9 * * *"))
def daily_post():
    post = generate_post.remote("AI automation trends")
    # Post to LinkedIn, etc.
    print(f"Generated: {post[:100]}...")

3. Add secrets

# Add secrets via Modal dashboard or CLI
modal secret create my-secrets \
  EURI_API_KEY=your-key \
  LINKEDIN_ACCESS_TOKEN=your-token

Then reference in code:

@app.function(secrets=[modal.Secret.from_name("my-secrets")])
def generate_post(topic: str):
    ...

4. Deploy

# Test locally first
modal run app.py

# Deploy to cloud (persistent)
modal deploy app.py

5. Call your endpoint

curl -X POST https://your-app--webhook.modal.run \
  -H "Content-Type: application/json" \
  -d '{"topic": "AI in 2026"}'

GPU Functions

@app.function(gpu="T4")  # or "A100", "H100"
def run_inference(prompt: str):
    # GPU-accelerated inference
    ...

GPU pricing:

• T4: $0.000164/sec (~$0.59/hr)
• A100: $0.00044/sec (~$1.58/hr)
• H100: $0.001097/sec (~$3.95/hr)

$30 free credit = ~50 hours of T4 time.

Gotchas

1. Cold starts — first call after idle takes 2-5 seconds. Use keep_warm=1 for critical endpoints.
2. $30 credit burns fast on GPU — use CPU for non-ML tasks.
3. Secrets are separate — not read from .env. Use Modal dashboard or modal secret create.
4. Python only — no Node.js, no Docker (Modal has its own container system).
5. modal deploy vs modal run — run is one-time test, deploy is persistent.

---

Schema

Inputs

Name	Type	Required	Description
script	file_path	Yes	Python file to deploy
gpu	string	No	GPU type: T4, A100, H100 (default: none/CPU)

Outputs

Name	Type	Description
url	string	Webhook endpoint URL
deployed	boolean	Whether deployment succeeded

Credentials

Name	Source
MODAL_TOKEN_ID	modal setup (automatic)
MODAL_TOKEN_SECRET	modal setup (automatic)
App secrets	Modal Dashboard or modal secret create

Composable With

vercel-deploy (Vercel frontend → calls Modal backend), github-actions-deploy (cron triggers Modal)

Cost

Free ($30/mo credit), then pay-per-second compute