Skip to main content
Connect mcp-agent to Magic Hour MCP for image generation and recovery. This command-line example uses the framework’s native tool calls without a language model or another provider key.
Verified October 24, 2026 with mcp-agent 0.2.6 / MCP 1.30.0: 48 discovered tools, an authenticated account check, one accepted image creation, same-ID status retrieval, and native inline image recovery produced a usable 640×640 PNG. Full image decode passed. This verification used native tool calls without a language model.

​
Install and configure

Requires Python 3.11+; native validation used Python 3.12. Copy magic_hour.py and requirements.txt into an empty directory, then run:
Provide your Magic Hour API key through your secret manager as MAGIC_HOUR_API_KEY. The MCP endpoint is fixed to https://mcp.magichour.ai/; the example disables logging and OpenTelemetry and does not print account fields, credentials, prompts, or signed download URLs.
The authorization command prints only its status, discovered tool count, and action count. Public tool discovery alone does not prove the API key works.

​
Create one image

The create command sends your prompt to Magic Hour and spends credits. It requests one square 640px Flux 2 Klein image, starting at 5 credits. Run it once for each intended image:
The example calls the native generation tool once and writes the accepted project ID to image-project.txt before returning. It does not poll or automatically repeat creation. The exclusive ID file prevents accidentally submitting the same command twice. If the response is interrupted, the file retains an unknown-outcome marker. Check your Magic Hour projects before submitting another creation. A timeout does not prove the job failed. If saving the accepted ID fails, the ID is still printed so you can preserve it.

​
Retrieve and download

Use the saved ID for all subsequent reads:
get prints the ID and current status. download waits up to 30 seconds for the saved job and saves the first native inline image, up to 8 MiB. Larger outputs require a separate exact-URL download. These commands reuse the same project. An unfinished job or failed download never triggers a new generation. Keep the saved ID and retry recovery after the job completes. An existing output file is not overwritten.

​
Use in your application

The example uses MCPApp and a native Agent configured for streamable_http. It discovers namespaced tools such as magic_hour_account_retrieve and invokes Agent.call_tool directly. The asyncio executor calls each requested operation once; usage telemetry, tracing, file logging, and automatic agent discovery are disabled. Both application and agent contexts close on exit. For a model-backed agent, expose only the tools your workflow needs and require explicit approval before paid creation. The native tool recipe does not establish model-driven behavior. See the framework source and our privacy policy. Tool errors are reported without printing private provider responses.