Connect LM Studio (local models)

LM Studio runs open models such as Qwen or Llama on your own computer. Because the model runs on your computer, the AI's work on your site does not go through a cloud AI service.

Before you start#

• LM Studio 0.3.17 or newer, and a model that supports tool use.
• A connection for LM Studio. Open PowerWP AI › Connections, choose Create connection, pick LM Studio, name it and choose what it may do. Choose Create key and show steps: the window shows the steps below with your address and key already in them. The key is shown once. Here it appears as <your key>.

Option 1: one click#

1. In the setup window, click Add to LM Studio.
2. LM Studio opens and asks you to install your site's server. Confirm it.
3. Load a model with tool support and start a chat. LM Studio may ask you to allow each tool call.

Option 2: add it by hand#

1. In LM Studio, open the Program tab in the right sidebar, then Install › Edit mcp.json.
2. Paste the JSON below and save. If the file already lists other servers, paste only the your-store entry inside mcpServers, after the last entry, with a comma between the two.
3. Load a model with tool support and start a chat.

{
  "mcpServers": {
    "your-store": {
      "url": "https://your-store.com/wp-json/wpai/v1/mcp",
      "headers": { "X-WPAI-Token": "<your key>" }
    }
  }
}

Check it works#

Name the tool so the model knows what to do:

Use the site_overview tool and summarize my site.

The model answers with your WordPress version, theme and plugins.

Troubleshooting#

• The model answers without using your site. Pick a model that supports tool use, and ask for one tool by name.
• Your site is not listed. Check your LM Studio version (0.3.17 or newer) and that mcp.json is valid JSON.
• 401 "Authentication required". The key was copied incompletely. Keys cannot be shown twice: revoke the connection and create a new one.
• 403 or 429. 403 means your IP address is not on the connection's allow-list, or a firewall blocks /wp-json/wpai/v1/. 429 means too many requests: wait a few minutes.

Related#