Models and Keys

Start with local memory mode unless you know you need automated language-heavy maintenance. It gives you storage, embeddings, dedupe, hybrid search, MCP recall, and local artifacts.

For background organization or page synthesis, configure an AI service in Settings and assign it to the tasks you want. Choose on-device Qwen, an OpenAI-compatible local server such as Ollama or LM Studio, or a supported cloud provider.

1. Open Settings > Intelligence. Under AI services, choose an on-device model, a local server, or a cloud provider. Complete the setup shown on its card. For cloud providers, enter your API key and Save; if the card offers Test, use it before saving.

2. Choose Automatic under Steeping, then assign a configured service to Everyday work and/or Synthesis. A cloud service sends the material needed for its assigned task to that provider.

3. Run wenlan steep status to check whether each task is ready, paused, or off. If a task is not ready, check its service and assignment, then check Activity for progress and the page's review queue for proposed updates.

Local memory mode uses the local search embedding model without requiring an additional language model. It supports brief, capture, recall, handoff, agent-side distill, and inspection of local artifacts.

This mode is also the right starting point when you want the smallest install path or when a machine cannot run local model inference comfortably.

An on-device model lets the daemon perform more language work locally. It can help with classification, extraction, title generation, recaps, and other maintenance tasks without sending those daemon-side prompts to an external API.

Local model capability is still bounded by hardware and context limits. For large distillation jobs, an API provider can be stronger when you choose to configure one.

An OpenAI-compatible endpoint on your machine can supply a model locally. The app also offers cloud provider presets, including Anthropic, OpenAI, Google, Groq, OpenRouter, Mistral, DeepSeek, and xAI. Configure the service in Settings and assign it to background organization or page synthesis.

When you choose a cloud provider for enrichment, Wenlan sends the material being processed and its prompt to that provider. Review its terms before configuring a key. Your connected AI client may separately send content to its own provider.

Claude Code skills can classify captures, write handoffs, and organize pages from inside the active agent session. That is why Wenlan remains useful before you configure a daemon model.

The daemon stays the store and retrieval layer. The agent can supply language judgment when it is already in the work context.

In the desktop app this is the On demand steeping mode, the default for new installs: your AI tool steeps as it saves or writes, and /wenlan steep finishes the rest. Automatic hands the same work to the models you assign here, in the background.