Build an agent

Model APIs call remote MCP servers for you: pass Argil’s URL and an API key, and the model uses Argil’s tools within one request. Agent frameworks connect as MCP clients in your process. Create a dedicated API key per agent in Workspace settings → API keys, so you can revoke one without touching the others.

Anthropic API
OpenAI API
xAI API
OpenAI Agents SDK
Vercel AI SDK

import os, anthropic

client = anthropic.Anthropic()
message = client.beta.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=4096,
    betas=["mcp-client-2025-11-20"],
    mcp_servers=[{
        "type": "url",
        "url": "https://mcp.argil.ai/mcp",
        "name": "argil",
        "authorization_token": os.environ["ARGIL_API_KEY"],
    }],
    tools=[{"type": "mcp_toolset", "mcp_server_name": "argil"}],
    messages=[{"role": "user", "content": "Make a square product shot of a glass perfume bottle on black velvet. Spend at most 5 credits."}],
)

import os
from openai import OpenAI

client = OpenAI()
response = client.responses.create(
    model="gpt-6",
    tools=[{
        "type": "mcp",
        "server_label": "argil",
        "server_url": "https://mcp.argil.ai/mcp",
        "authorization": os.environ["ARGIL_API_KEY"],
        "require_approval": "never",
    }],
    input="Make a square product shot of a glass perfume bottle on black velvet. Spend at most 5 credits.",
)

require_approval: "never" runs tools without a round trip to your code. Keep a budget in the prompt, or add your own approval step.

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["XAI_API_KEY"], base_url="https://api.x.ai/v1")
response = client.responses.create(
    model="grok-4.7",
    tools=[{
        "type": "mcp",
        "server_label": "argil",
        "server_url": "https://mcp.argil.ai/mcp",
        "authorization": os.environ["ARGIL_API_KEY"],
    }],
    input="Make a square product shot of a glass perfume bottle on black velvet. Spend at most 5 credits.",
)

The xAI API runs tools without asking: keep a budget in the prompt.

import os
from agents import Agent, Runner
from agents.mcp import MCPServerStreamableHttp

async with MCPServerStreamableHttp(
    name="argil",
    params={
        "url": "https://mcp.argil.ai/mcp",
        "headers": {"Authorization": f"Bearer {os.environ['ARGIL_API_KEY']}"},
        "timeout": 60,
    },
    cache_tools_list=True,
) as argil:
    agent = Agent(name="Creative", instructions="You make ad creatives with Argil.", mcp_servers=[argil])
    result = await Runner.run(agent, "Three 9:16 hooks for a running shoe, 5 seconds each, at most 25 credits each.")

import { createMCPClient } from "@ai-sdk/mcp";
import { generateText, stepCountIs } from "ai";
import { anthropic } from "@ai-sdk/anthropic";

const argil = await createMCPClient({
  transport: {
    type: "http",
    url: "https://mcp.argil.ai/mcp",
    headers: { Authorization: `Bearer ${process.env.ARGIL_API_KEY}` },
  },
});

const { text } = await generateText({
  model: anthropic("claude-sonnet-5-5"),
  tools: await argil.tools(),
  stopWhen: stepCountIs(12),
  prompt: "Make a square product shot of a glass perfume bottle on black velvet. Spend at most 5 credits.",
});
await argil.close();

Design rules for a media agent

Send an idempotency key. Have the agent pass a new UUID as idempotency_key for each generation it means to make, and the same one when it retries. A retry then returns the generation already started instead of paying for a second one.
Give it a budget. “At most N credits each” becomes max_credits: Argil refuses anything above it, before spending.
Let it wait, in steps. A call returns within about 40 seconds. A video comes back as processing, and the agent checks again with get_generation. Allow enough steps: a 4-minute video takes about 6 checks.
Tag what it makes. metadata (for example {"sku": "A-102"}) lets it find generations later with list_generations.
Don’t make the model wait on a batch. For many generations, start them all, then let a webhook endpoint (create_webhook_endpoint) tell your server when each one finishes.

See How it works for the details behind these rules.

Get started

Guides

Reference

Build an agent

Design rules for a media agent

​Design rules for a media agent

Design rules for a media agent