# AI coding CLIs GUIJam can drive

> GUIJam drives the AI coding CLI you already use and are logged into, such as Claude Code, Codex or OpenCode. No API keys to set up.

GUIJam doesn't have its own model or API key. It drives the AI coding CLI you already use and are logged into, headlessly, and asks it for structured edits.

| AI tool                                                                  | ID            | Images and sketches  | Verified live |
| ------------------------------------------------------------------------ | ------------- | -------------------- | ------------- |
| [Claude Code](https://guijam.francistinao.com/docs/ai-tools/claude-code) | `claude-code` | Yes                  | Yes           |
| [Codex](https://guijam.francistinao.com/docs/ai-tools/codex)             | `codex`       | Yes                  | Yes           |
| [OpenCode](https://guijam.francistinao.com/docs/ai-tools/opencode)       | `opencode`    | Depends on the model | Yes           |
| [Qwen Code](https://guijam.francistinao.com/docs/ai-tools/qwen-code)     | `qwen-code`   | No                   | Yes           |
| [Gemini CLI](https://guijam.francistinao.com/docs/ai-tools/gemini-cli)   | `gemini-cli`  | No                   | Yes           |

## Choose a tool

`init` lists the tools on your machine and saves the one you pick. To see them again, or switch:

```bash
npx guijam provider list
npx guijam provider use codex
```

The choice is saved for your user, in `~/.config/guijam/config.toml` on macOS and Linux and `%APPDATA%\guijam\config.toml` on Windows, and applies to every project. GUIJam reads it on every request, so a switch takes effect on your next request without restarting the dev server.

When you haven't picked one, GUIJam uses the first tool that is logged in, in the order of the table above. With none logged in, it falls back to Claude Code.

## Choose a model

GUIJam uses a fast model for quick style changes and a standard model for larger ones:

| AI tool             | Fast requests            | Larger requests          |
| ------------------- | ------------------------ | ------------------------ |
| Claude Code         | `haiku`                  | `sonnet`                 |
| Codex               | `gpt-6-luna`             | Your Codex default       |
| Gemini CLI          | `gemini-3-flash-preview` | `gemini-3-flash-preview` |
| OpenCode, Qwen Code | Your CLI's default       | Your CLI's default       |

Pick a model for any tool with `--model`. It's used for every request:

```bash
npx guijam provider use opencode --model deepseek/deepseek-chat
```

## Choose an effort

GUIJam sets how hard Claude Code and Codex think on each request, instead of using your CLI's own setting: low effort for small edits and questions, medium for sketches, images, links and open-ended changes. To use one level for every request, pass `--effort` with `low`, `medium` or `high`:

```bash
npx guijam provider use claude-code --effort high
```

Higher effort can give better results on larger changes, but takes longer and uses more of your plan. OpenCode, Qwen Code and Gemini CLI ignore the setting.

## Logging in

GUIJam never reads, copies or stores your credentials. It checks your login with the CLI's own status command, and `init` runs the CLI's own login command if you aren't logged in. Login status is cached for 5 minutes.

A login can look fine and still fail on a real request, for example with an expired key or an account out of credit. For OpenCode, Qwen Code and Gemini CLI, `init` offers to send one short test request, which spends a few tokens. To check any tool at any time:

```bash
npx guijam doctor --probe
```

## When a request fails

The widget repeats what your CLI said, so you see the real reason, such as "OpenCode was refused: Invalid API key." or "Gemini CLI stopped before answering: `<the last line it printed>`." It also tells you what to do next. Every failed run is logged in `.guijam/logs/gjd.log`, with the tool, its version and the model, and with account IDs replaced by `redacted`. See [Troubleshooting](https://guijam.francistinao.com/docs/troubleshooting#ai-tool-problems) for each message.

## What your CLI can and can't do

Every run is headless, and your CLI gets no tools that can change files. GUIJam passes the relevant code inline, and the CLI answers with edits as data. The `gjd` daemon validates those edits and is the only process that writes to your project. See [Security and privacy](https://guijam.francistinao.com/docs/concepts/security).
