Skip to main content
Connect AnythingLLM’s native MCP client to Magic Hour, then use @agent to retrieve an image project you already created. This setup enables only the read-only image retrieval tool. It does not require another generation or a new model-provider key when you use your own local Ollama model.

​
Before you start

Use a protected self-hosted AnythingLLM instance with administrator access, a Magic Hour API key, and an existing image project owned by that account. These instructions cover self-hosted AnythingLLM 1.18.0. Have a tool-capable model installed and reachable from AnythingLLM. This flow was verified with Ollama 0.40.10 and qwen2.5:7b-instruct-q4_K_M at a 4096-token context: one native agent run retrieved an existing completed image project. No new Magic Hour generation was needed. Other models may choose or format tool calls differently.
AnythingLLM stores MCP headers as plaintext in its instance-wide configuration file. Administrators and anyone who can read that file can access the API key; other enabled Magic Hour tools could spend credits. Use a dedicated key, restrict the file to the application owner with permissions 0600, protect the instance, and keep configuration exports, logs and screenshots private. Do not put the key in a chat prompt.

​
Discover and restrict the tools

  1. In the instance’s storage directory, open or create plugins/anythingllm_mcp_servers.json. Merge this entry into any existing mcpServers object; preserve other servers:
    The full path is <STORAGE_DIR>/plugins/anythingllm_mcp_servers.json. The standard Docker storage path is /app/server/storage/plugins/anythingllm_mcp_servers.json; edit the corresponding file in the host storage mount.
  2. Open Settings → Agent Skills. In the MCP section, click Refresh and confirm the server restart. Select Magic Hour. Public discovery returns tools without an API key; it does not prove account access.
  3. Turn off every Magic Hour tool except image_projects_retrieve_details. The verified inventory had 48 tools, leaving 1/48 tools enabled.
  4. Turn off RAG & long-term memory, View & summarize documents, Scrape websites, and Web Search, then click Save. Keep other agent skills, imported skills, flows and clarifying questions disabled for this read-only setup.
  5. Reload the page and refresh the MCP server once more. Confirm that only the image retrieval tool remains enabled and the four core skills remain off.
Tool selection and MCP configuration apply across the instance. Use a dedicated instance if changing these settings would affect other users or workspaces.

​
Add the API key

Replace the empty headers object with:
Replace the placeholder with your actual key in the protected configuration file. AnythingLLM reads this as a literal header: ${MAGIC_HOUR_API_KEY} is not expanded from an environment variable. Keep the Bearer prefix and never commit the file. Refresh the MCP server through the same native UI, then check that the one-tool selection is still saved. A connected server or successful discovery alone does not verify the API key; the authenticated project request does.

​
Configure the model and workspace

  1. Open AI Providers → LLM and select Ollama. Set its reachable base URL, select your installed model, set the context size, and click Save changes. If AnythingLLM runs in Docker, localhost refers to that container; use the Ollama service’s reachable address on your configured network.
  2. Create a workspace. In its Agent Configuration, select Ollama and the same model. Allow the normal settings save to finish.
  3. Send one message in that workspace, replacing YOUR_PROJECT_ID with your own saved image project ID:
The request retrieves a status snapshot; it does not poll, create a replacement job or download an image. If the project is still processing, check the same ID again later. Your model’s inference may have costs even though this Magic Hour operation creates no new generation.

​
Check the result

Look for the native tool invocation and its completion before relying on the agent’s answer. Confirm that the returned ID matches your saved project and inspect its status. A model-written complete without an actual successful tool invocation is not evidence that Magic Hour was contacted. For an authorization error, check the literal bearer header and that the project belongs to the key’s account. After editing the configuration, refresh the native MCP server and recheck the saved tool selection. Do not enable creation or deletion tools to troubleshoot a read-only retrieval. Project results may contain temporary signed download URLs. Keep raw tool results and instance logs private; do not share them with the configuration file or key.