Use Miles with a model that runs on your own computer

What you end up with

• A model running on your laptop, connected to Miles, answering wallet questions.
• Nothing you type, attach, or read leaving the machine. The only traffic is the tool calls to Miles, each one shown to you first.
• Every other guide on this site available to you with this setup, unchanged.

Before you start

1. A computer that can run a small model. An Apple silicon Mac with 8 GB of memory is enough for a 4 billion parameter model; 16 GB runs the 8 to 20 billion class comfortably. A PC needs a graphics card with 6 GB or more of memory for the same.
2. About 5 GB of disk for one model.
3. A Miles account with your cards entered. The model reads what is there. Sign in on the web once so the login step below has an account to approve.
4. Read this page once: Your data and your AI says what a connected app sends Miles and what it never can. It applies to a local model exactly as it does to Claude.

What leaves your device at each step, and what never does: Your data and your AI.

Step 1: Install LM Studio and pick a model

LM Studio is the shortest path: one app that runs the model and connects to Miles, on Mac, Windows and Linux. Install it, open the model search, and choose a model that says it supports tool use. Not every small model does, and one that does not will read the tools and never call them.

Gemma 3 runs well on the same hardware and reads a document nicely, but check its listing for tool use before choosing it for this. Download the model and load it. In the model's settings, set the context length to at least 8,000 tokens: Miles offers a few dozen tools, and their definitions need room.

Step 2: Connect Miles

In LM Studio, open the Program panel on the right, then Install, then Edit mcp.json (the location as of version 0.4.10 and later, the first release with a browser login for remote servers). Add Miles:

Save. Your browser opens to Miles. Sign in, and on the approval screen decide whether to turn on Also allow changes to your wallet. Leave it off for now; reading is all this guide needs, and a read-only login is offered only the read tools, which is also easier on a small model. Approve, and LM Studio shows Miles in its connector list with the tools it received.

Step 3: Ask something

Enable the Miles connector in the chat and paste this:

LM Studio shows each tool call the model wants to make and waits for you to allow it. The first time, allow them one by one and read what each sends: a tool name and a small argument, nothing else. The answer comes back from your wallet, with the earning rates Miles used.

Then try one that needs a file. Attach a PDF you would not upload to a cloud service, a statement say, and follow the credits and free nights from your statements guide. The file is read on your machine. Miles receives a status per credit, and nothing from the PDF itself.

If it does not work

The browser never opens, or the login loops. Update LM Studio; the browser login arrived in 0.4.10 and was fixed for some Windows setups in 0.4.12. If you approved once and it still loops, open Miles Settings, Connected AI Apps, disconnect the LM Studio entry, and add the server again.

The model reads the tools and never calls one. Usually the model. Pick one listed with tool use, and raise the context length: a model whose context is full of tool definitions has no room to plan. If the model calls a tool with an empty or made-up argument, try the next size up.

Miles refuses a call. The message says why: a tool that needs Membership, a write without the toggle on, or a paused connection. Each is a setting on your side, not a fault in the model.

Other ways to run the model

• Open WebUI, with Ollama or any local server behind it: add https://mcp.askmiles.ai/mcp under Settings, External Tools, as type MCP (Streamable HTTP) with OAuth 2.1 (version 0.6.31 or later). The login runs in the browser the same way.
• Ollama on its own has no connector support built in. Put a client in front of it that uses it for the model: Goose, mcphost, or ollmcp all connect to remote MCP servers and can sign in.
• An agent you run yourself. OpenClaw connects the same way, and its model is your choice.

Miles has not tested every client. The connector is standard, and the login is the same one the big apps use, so a client that can sign in to any OAuth-protected MCP server can sign in to this one. If one refuses, say which.

What this does and does not buy you

Your conversation, your files, and everything the model reads or writes stay on your computer. Your wallet still lives with Miles, encrypted with a key that belongs to your account; the security page covers how. A local model removes the AI vendor from the picture. It does not remove Miles, which is where the wallet was all along.