Meta Model API cookbook

Recipes for building agents and coding tools on the Meta Model API, covering API fundamentals, agent patterns, and end-to-end use cases.

Meta Model API lets you build with Muse Spark using the tools you already run: it's drop-in compatible with the OpenAI SDK, the Anthropic SDK, and agent CLIs like OpenCode and Claude Code.

Point your client at the Model API base URL, set your key, and keep the rest of your code. Each recipe is a self-contained, copy-paste starting point that proves a capability and gives you something to build on. The default model is Muse Spark (muse-spark-1.1), which has a 1,048,576-token context window; the preview is free.

Getting started

You need a Model API account and an API key. Store the key as an environment variable so it stays out of your code:

export MODEL_API_KEY="LLM|{numeric_id}|{secret}"
pip install openai

import os

from openai import OpenAI

# The OpenAI SDK does not auto-read MODEL_API_KEY, so pass it explicitly.
client = OpenAI(
    base_url="https://api.meta.ai/v1",
    api_key=os.environ["MODEL_API_KEY"],
)

response = client.chat.completions.create(
    model="muse-spark-1.1",
    messages=[{"role": "user", "content": "Hello, world!"}],
)

print(response.choices[0].message.content)

Recipes

This cookbook mirrors the three sections of the Meta Model API cookbook website. Each recipe is a self-contained, copy-paste starting point. Start with API fundamentals to learn the API primitives, then move into agent patterns and end-to-end use cases.

1. API fundamentals

Prove each API primitive works and get a starting point you can build on.

#	Recipe	What it does
01	Quickstart: chat completions	Make your first Muse Spark call by pointing the OpenAI SDK at one new base URL.
02	Streaming responses	Render tokens as they generate and read the final usage chunk.
03	Tool and function calling	Detect tool calls and run the execute-and-feed-back loop.
04	Structured output	Get schema-guaranteed JSON that parses on the first try.
05	Prompt caching	Reuse a stable prompt prefix and track cached tokens.
06	Reasoning and thinking tokens	Control reasoning effort and replay reasoning across turns.
07	Vision input	Send images by URL or base64 and get structured analysis back.
08	Long context	Pack repo-scale context into the context window.
09	Error handling and retry	Back off with jitter and skip retries on client errors.
10	Search grounding	Ground responses in live web search results with inline citations.

2. Agent patterns

Build the loops that turn a model into an agent: planning, parallel work, and self-correction.

#	Recipe	What it does
01	Basic agent loop	Wire up the core perceive-decide-act agent loop.
02	Interleaved reasoning and tool use	Interleave reasoning with tool calls in a single turn.
03	Multi-turn context management	Manage growing context across a long agent run.
04	Validated in-place edits	Validated search-and-replace edits with a coding agent.
05	Alert fatigue copilot	Extract grounded patterns from a noisy alert feed, then probe → chat → self-assess with strict-JSON output.

3. Use cases

End-to-end patterns: multimodal perception, orchestration, and full applications.

#	Recipe	What it does
01	Chart analysis	Read charts and extract structured data from images.
02	Error screenshot fix	Diagnose a bug from an error screenshot and fix it.
03	Smart glasses with OpenClaw	Hands-free look-and-ask on Ray-Ban Meta glasses.
04	Generating slides	Generate a slide deck from a prompt or source content.
05	Browser-verified web design	Build a website with a coding agent that checks its own work in a real browser.
06	Iterative game dev	Build a browser game end-to-end with a coding agent, verified in a real browser.
07	Sandboxed execution	Execute model-generated code in a sandbox.
08	Multi-agent product studio	Orchestrate a four-profile Hermes team (PM, backend, frontend, tech writer) that coordinates through a shared Kanban board.
09	One-shot game dev	Build a complete 3D browser game in a single pass — an AGENTS.md plus one structured prompt, no iteration loop.
10	Perception grounding	Identify food items in a fridge photo, pin interactive health-score dots at their pixel locations, and generate a self-contained HTML overlay.
11	GitHub repo agent	Autonomous GitHub Actions bot — triage, PR review, AI-slop detection, and bug-fix PRs — built on OpenCode + Muse Spark.
12	Computer use	Drive a Linux desktop from screenshots — the agent finds an app, opens it, and plays it, clicking through a Cua sandbox.
13	macOS computer use	Drive a real Mac from screenshots with metacua, a native computer-use agent (Swift + Python) that clicks, types, and works out a GUI app on its own.

License

See LICENSE for details.