Repository navigation
Choosing a mode
A mode decides what Dive talks to: models on your own computer, the pi coding agent, or a cloud provider. You can switch at any time, and each mode keeps its own conversations and settings. This page compares the five so you can pick a starting point.
| Mode | Talks to | You need | Shown at first launch |
|---|---|---|---|
| llama.cpp |
llama-server from llama.cpp |
llama.cpp and a GGUF model file. Dive's MODELS tab can download models from Hugging Face and run llama-server for you |
Yes |
| Pi | The pi coding agent, which Dive starts as pi --mode rpc
|
pi installed on your computer | Yes |
| Cloud | OpenAI, Anthropic, Mistral or Google Gemini | An API key for the provider | Yes |
| LM Studio | LM Studio's local server, http://127.0.0.1:1234/v1 by default |
LM Studio with its server running | No |
| Ollama | Ollama's API, http://127.0.0.1:11434 by default |
Ollama running | No |
Each mode has its own tutorial: llama.cpp, Pi, Cloud, LM Studio and Ollama.
| llama.cpp | LM Studio | Ollama | Cloud | Pi | |
|---|---|---|---|---|---|
| Dive's native, external and custom tools | Yes | Yes | Yes | Yes | No, pi has its own |
| MCP servers | Yes | Yes | Yes | Not shown in Settings | No |
| Agent Skills | Yes | Yes | Yes | Yes | No, pi loads its own |
| NATIVE TOOL CALLING setting | Yes | Yes | Yes | No, tools are described in the prompt | Not applicable |
| Agent Mode | Yes | Yes | Yes | Yes | Not applicable |
| Tool calls per request, without Agent Mode | 6 | 6 | 10 | 6 | Decided by pi |
| Runs on your computer | Yes | Yes | Yes | No, requests go to the provider | Depends on the model pi uses |
Agent Mode raises the tool budget to 25 calls per request by default, adjustable from 1 to 50, and asks the model to plan before it works. How the model uses tools explains both settings.
- New to local models? Start with llama.cpp. Dive can download a model and run the server from its MODELS tab, so you don't need the terminal.
- Already using LM Studio or Ollama? Use that mode and keep managing your models where you do now.
- Want the largest models? Use Cloud. Your messages go to the provider you choose, together with passages from your library when the DATABASE switch is on.
- Want an agent that works on your files and runs commands? Use Pi. Dive shows pi's work and asks you about its permission requests.
Click a mode's icon in the top-left switcher. To choose which icons appear, open Settings, go to the MODES tab and tick them under ENABLED MODES. At least one mode has to stay on. DEFAULT MODE ON LAUNCH picks the mode Dive opens in; "First enabled" means the leftmost one.
Every mode has its own conversation history, chosen model, colour palette and font.
The four Dive modes (llama.cpp, LM Studio, Ollama and Cloud) also keep their own:
- system prompt and lessons
- tool and skill switches, and custom tools
- MCP servers
So you can, for example, keep shell_command on in llama.cpp and off in Cloud, or give each mode a different system prompt. In Pi mode, pi manages its own tools, skills and context files.
Modes are defined once, in assets/js/00-modes.js, which both the browser and the server load. The diveSkills flag is false only for Pi, and localEndpoint marks llama.cpp and LM Studio, which share /api/<mode>/stream. The enabled modes and the default mode are saved with the interface settings through /api/ui/settings. Architecture has the details.
These pages describe Dive 5.0.8. Source code: DuckTapeKiller/dive. Spotted a mistake? Open an issue.
Get started
Tools and skills
Build your own
Share and stay safe
How Dive works
Help