# Set up Magnitude with an uncensored model

You are an AI coding agent. The person who gave you this file wants you to set up
[Magnitude](https://magnitude.dev), a local inference engine, and pull the best uncensored model
that fits their hardware, then connect their agent to it.

Do the setup end to end: install Magnitude, pick and pull the right model, and connect the harness.
When it is wired up, tell the owner how to run their agent on the new model, and stop.

## 1. Install Magnitude

macOS / Linux:

```sh
curl -fsSL https://magnitude.dev/install.sh | sh
```

Windows (PowerShell):

```powershell
irm https://magnitude.dev/install.ps1 | iex
```

This installs the desktop app and the `magnitude` CLI together. On macOS, the `magnitude` command
is available in a new shell, so start one before continuing.

Then make sure a service is running:

```sh
magnitude status
```

If it reports no running service, start one in the background and leave it running:

```sh
magnitude serve &
```

## 2. Learn the model commands

Read the custom-models guide before pulling anything. It explains how Magnitude adds models from
Hugging Face and how to read a fit verdict for this machine:

```sh
magnitude docs custom-models
```

## 3. Pick a model that fits this machine

Check the hardware first:

```sh
magnitude hardware
```

Then inspect these uncensored / abliterated models, largest first, and pull the **largest one that
fits**. `catalog inspect` prints a `Fits`, `Too large`, or `Unsupported` verdict for this machine
and changes nothing, so run it on all of them before deciding. A `Fits` verdict already accounts for
the memory the OS and other apps need, so trust it rather than holding back extra.

```sh
magnitude catalog inspect orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF
magnitude catalog inspect HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
magnitude catalog inspect HauhauCS/Qwen3.5-9B-Uncensored-HauhauCS-Aggressive
magnitude catalog inspect HauhauCS/Qwen3.5-4B-Uncensored-HauhauCS-Aggressive
```

Pull the one you chose. With no quant, Magnitude picks the recommended one for this machine:

```sh
magnitude catalog pull <owner/repo>
```

`catalog pull` prints the exact **model ID**. Keep it; you need it for the next step. If a repo is
gone or refused, fall back to the next smaller one in the list.

## 4. Connect the harness to the local model

Figure out which agent the owner uses, then connect that harness and point it at the model you
pulled. Replace `<harness>` with the right ID from this list:
`pi`, `opencode`, `hermes`, `openclaw`, `codex`, `claude-code`, `oh-my-pi`, `cline`.

```sh
magnitude connections connect <harness> --set-model <model-id>
```

This writes the configuration. It does **not** switch the model of any session already running, so
the owner runs their agent on the local model in a new session.

## 5. Tell the owner how to run on the local model

Setup is done. Tell the owner which model you pulled, and give them the command that starts their
agent on it from the root of this project. Fill in the `<model-id>` you pulled.

| Agent | Command to start on the local model |
| --- | --- |
| Claude Code | `claude --model anthropic-local/<model-id>` |
| Codex | `codex -c model_provider=magnitude --model magnitude-local/<model-id>` |
| Pi | `pi --provider magnitude --model <model-id>` |
| Oh My Pi | `omp --model magnitude/<model-id>` |
| Hermes | `hermes --provider custom:magnitude --model <model-id>` |
| Cline | `cline --tui --provider openai-compatible --model <model-id>` |
| OpenClaw | `openclaw tui`, then pick the model in its model selector |
| OpenCode 1 | `opencode --model magnitude/<model-id>` |
| OpenCode 2 | `OPENCODE_CONFIG_CONTENT='{"model":"magnitude/<model-id>"}' opencode --standalone` |

OpenCode 2 dropped `--model` and shares one background server, so check the installed version with
`opencode --version` and use the matching row. On Windows (PowerShell) the OpenCode 2 form is
`$env:OPENCODE_CONFIG_CONTENT='{"model":"magnitude/<model-id>"}'; opencode --standalone`.

For any other OpenAI-compatible agent, point it at base URL
`http://127.0.0.1:10100/inference/v1` with the model ID above.

Tell the owner the model runs entirely on their machine, so nothing they do with it leaves the
computer.
