Skip to content

Ollama mode

DuckTapeKiller edited this page Sep 13, 2026 · 2 revisions

Ollama mode talks to Ollama. Use it if you already pull and run models with Ollama: Dive lists your models, sends your messages through Ollama's chat API, and adds its tools and skills on top.

Before you start

  • Ollama installed and running. It listens on port 11434 unless you change it.

  • At least one model, pulled with Ollama, for example:

    ollama pull <model>

1. Switch the mode on

Ollama mode is off at first. Open Settings, go to the MODES tab and tick Ollama under ENABLED MODES. Its icon appears in the top-left mode switcher. Click it.

2. Connect to Ollama

In Settings, on the MAIN tab, check OLLAMA SERVER URL. The default is http://127.0.0.1:11434. You can point it at Ollama on another computer: chatting, listing models and model information all use this address.

3. Pick a model

Press Refresh models, then choose a model from MODEL. The choice stays in step with the MODEL dropdown in the side panel. After you pull a new model, or change the server address, press Refresh models again.

4. Start chatting

Close Settings and type a message. Replies stream in from Ollama.

5. Try a tool and a skill

Ask something that needs a lookup:

What time is it in Tokyo right now?

The model calls the time_and_date tool and answers from the result. /calc 12 * 44 runs a tool directly.

Most native tools start switched on. shell_command, macos_control and propose_plugin start switched off; turn them on in Settings > TOOLS. browse_act also starts off, but in 5.0.8 Settings has no switch for it or for browse_read. Tools that run code or act on a web page ask for your permission every time.

Type /skill:name and your text to use a skill from ~/dive/skills or ~/.agents/skills. See Create a skill.

Settings for Ollama mode

On the MAIN tab, in Ollama mode:

Setting What it does Default
OLLAMA SERVER URL The address of the Ollama server http://127.0.0.1:11434
MODEL The Ollama model to use. Refresh models reloads the list
NATIVE TOOL CALLING (OPENAI SCHEMA) Sends tools in Ollama's tools field, which models trained for tool use follow best. The text prompt stays in the conversation for models that ignore tools. Off sends only MCP tools that way and everything else through the text prompt On
AGENT MODE Plan-first prompting and a larger tool budget. Tool calls appear as a step timeline in the chat Off
AGENT TOOL BUDGET Tool calls per request in Agent Mode, from 1 to 50. Without Agent Mode the limit is 10 25
COLOR PALETTE, FONT FAMILY, FONT SIZE How this mode looks

OLLAMA GENERATION

These options go to Ollama with every request. Press SAVE OLLAMA OPTIONS after changing them, or RESET OLLAMA OPTIONS to go back to the defaults.

Option Sent as What it does Default
TEMPERATURE temperature Randomness: lower is more predictable 0.3
TOP P top_p Limits token choices to a cumulative probability 0.75
TOP K top_k Caps the pool of candidate tokens at each step 40
NUM PREDICT num_predict The most tokens to generate. -1 means no cap 2,048
NUM CTX num_ctx The context window in tokens 32,768
SEED seed The same prompt and seed help repeat an output -1
REPEAT PENALTY repeat_penalty Discourages repeating text 1.1
REPEAT LAST N repeat_last_n How many recent tokens the repeat penalty checks. -1 uses the full context 256
STOP SEQUENCES stop Strings that end the reply, separated by || or new lines None

How it works

For developers and agents:

  • The window posts to /api/chat/stream, handled in routes/chat.js, which calls Ollama's /api/chat and streams NDJSON.
  • The model list comes from Ollama's /api/tags.
  • With NATIVE TOOL CALLING on, getLocalNativeTools("ollama", …) fills the request's tools field. With it off, only MCP tools are sent that way.
  • In Ollama mode the window builds the tool prompt itself (getOllamaBasePolicyPrompt() in assets/js/01-core.js), and the generation options come from getOllamaOptionsRequestPayload() in assets/js/03-theme.js.

Troubleshooting

Symptom What to do
No models appear Check that Ollama is running and that OLLAMA SERVER URL is right, then press Refresh models
A model you just pulled is missing Press Refresh models
Replies get cut off Raise NUM PREDICT, or set it to -1
The model forgets earlier parts of a long conversation Raise NUM CTX, within what your model and memory allow
The model never calls tools Check NATIVE TOOL CALLING. Some models follow the text prompt better, so try switching it off

Read next

Clone this wiki locally