Local models (Ollama)

Ollama lets you run language models entirely on your own machine — requests never leave your device and no external API key is required. Kodik automatically discovers a running Ollama server and offers all pulled models for selection.

Enabling Ollama

Use Enable Ollama in Settings to control whether local models are available to chat. Turning it off removes Ollama models from the model selector immediately; turning it back on restores the models Kodik has already discovered without requiring a reload.

Installing Ollama

If Ollama is not yet installed, Kodik will offer to install it from the UI. You can also install it manually.

OS	Command
Linux	curl -fsSL https://ollama.com/install.sh | sh
macOS	brew install ollama
Windows	winget install -e --id Ollama.Ollama

Windows and PATH

Kodik captures PATH at launch time, so installing Ollama after opening the IDE may go unnoticed — ollama appears “not found” even though it works in a freshly opened terminal. Kodik handles this by looking for the binary directly in its known per-user location (%LOCALAPPDATA%\Programs\Ollama\ollama.exe and similar) without relying on PATH. If you still see issues, restart Kodik.

Starting and stopping the server

Kodik can start and stop the Ollama server from the built-in UI:

• Start — Kodik runs ollama serve in the background. The process stays alive even after the terminal window closes.
• Stop — Kodik sends the appropriate shutdown command for the current OS.

You can also start the server yourself with ollama serve in any terminal.

Configuring the base URL

By default Kodik connects to http://localhost:11434. If you run Ollama on a different host or port — for example inside Docker or on a remote machine — enter the appropriate URL in the Ollama provider settings.

For Docker, remote, or proxied Ollama endpoints, Kodik uses the configured HTTP endpoint for model discovery and API-backed management actions. Model installation from Settings runs the local ollama command, so the CLI must be installed on the same computer as Kodik even when the configured endpoint is remote.

Request timeout

The Request timeout field in the Ollama provider settings sets how long, in milliseconds, Kodik waits for the model to start responding before it gives up on the request. A large model on a cold load, or a long prompt evaluated on CPU, can take longer than the default to produce its first token — raise the timeout so Kodik waits as long as you need. The default is 60000 (60 seconds); the value is in milliseconds.

Pulling and deleting models

The Ollama settings section offers two model-installation paths:

• Install on a recommended model — opens a visible integrated terminal and runs ollama pull <model>.
• Install by model name — opens the same terminal flow with the model ID you entered, for example ollama pull qwen3.5.
• Delete — removes the model from disk via the Ollama API.

The terminal shows Ollama’s native download progress and errors. Both paths target the configured Ollama base URL through OLLAMA_HOST.

If the model you need is not recommended, find its name on ollama.com/library and enter it under Install by model name.

Testing a model

The Test model button sends a short test message to the selected model and shows the response and response time. This is useful for verifying your setup before starting real work.

Context window discovery

Kodik discovers installed models through Ollama’s /api/tags endpoint and reads each model architecture’s supported context from /api/show. That model metadata is the source of truth shown in the context menu and model settings, and it also drives request budgeting and automatic compaction.

Kodik also checks /api/ps for running-model diagnostics, but a smaller runner allocation never replaces a context size successfully loaded from /api/show. For example, a model that advertises 40 960 tokens continues to display and use 40 960 even when /api/ps reports 4 096.

Kodik also reads /api/show model capabilities. When the selected model advertises tools, Kodik sends the active Harness tool schemas and tool guidance. When it explicitly omits that capability, Kodik sends a tool-free chat request instead. Goal authoring falls back to validated JSON, and a foreground sub-agent selected on that model can still complete a prose-only delegated task.

The Context window field in Kodik configures native Ollama features such as autocomplete; chat discovers the selected model’s context automatically. To change how Ollama allocates memory at runtime, change the context setting in Ollama itself, set OLLAMA_CONTEXT_LENGTH for the Ollama server, or create a model with PARAMETER num_ctx in its Modelfile.

Reasoning (thinking) control

Models that advertise the thinking capability (for example the Qwen3 family) get a per-model Reasoning toggle in the chat model picker. On keeps Ollama’s default behavior — the model thinks before answering and the thought process is shown in a collapsible block. Off sends think: false, so the model answers directly without a thinking pass — faster and cheaper on tokens for simple tasks. Models without the capability never receive the flag.

Authentication

If your Ollama server requires a Bearer token — for example when using Ollama Cloud or a private corporate instance — enter it in the API Key field in the Ollama provider settings. Leave it empty for a standard local installation.