FR
Agents

Ollama

Use Ollama with Palabre to bring in local models or models compatible with your Ollama configuration.

Ollama runs models on your machine. With a local server and model, requests stay local; a remote server receives the data. The first response may take time while loading: ollama ps lists loaded models.

Install before Palabre

Install Ollama from the official documentation, then download a model.

Official documentation: https://docs.ollama.com/

CLI reference: https://docs.ollama.com/cli

Then verify that the command works:

ollama list

or:

ollama ls

If Ollama was installed after your first Palabre configuration, open palabre again or synchronize explicitly:

palabre config --sync-agents

If needed, start Ollama:

ollama serve

Installing a model

ollama pull gemma4:e4b

Choose a model appropriate for your machine. Large models may be too heavy for smooth use.

In local use, the main limits come from your machine: memory, CPU/GPU, and response time. If you use a cloud, paid, or remote Ollama offering, quotas and limits come from that offering, not from Palabre.

Choose the model used by Palabre

Palabre reads the model from the ollama-local agent in your configuration. To list installed models and check whether the configured model still exists:

palabre config --ollama-models --json

To change Palabre's default Ollama model:

palabre config --set-ollama-model gemma4:e4b

If you removed the configured model and want Palabre to pick an installed model that is available:

palabre config --sync-ollama-model

These commands update the Palabre configuration. For a one-off debate override, use --model-a, --model-b, or --summary-model depending on where the Ollama agent is used.

Typical configuration

Excerpt from the agents block:

"ollama-local": {
  "type": "ollama",
  "baseUrl": "http://localhost:11434",
  "model": "gemma4:e4b",
  "role": "critic",
  "temperature": 0.2,
  "validateModel": true,
  "unloadOtherModels": true
}

Use a remote Ollama server

For one session, override the address of every Ollama agent:

palabre codex-ollama "Critique this plan" --ollama-url gpu-box:11434

You can also define OLLAMA_HOST in the process environment:

$env:OLLAMA_HOST = "gpu-box:11434"
palabre codex-ollama "Critique this plan"

Palabre accepts addresses with or without an HTTP scheme. The priority is:

  1. --ollama-url
  2. OLLAMA_HOST
  3. baseUrl in the agent configuration
  4. http://localhost:11434

From the TUI, open /config, then use:

/ollama-url gpu-box:11434
/ollama-url default

This command persistently updates baseUrl for every configured Ollama agent. The flag and environment variable do not modify the configuration file.

Automatically download a missing model

By default, Palabre does not install a model automatically. To allow it at startup:

palabre codex-ollama "Critique this plan" --pull-models

You can also enable autoPullModel in the agent configuration.

Give context to Ollama

Ollama does not read your files directly. Use:

palabre codex-ollama "Critique this module" --files src/module.ts

or:

palabre codex-ollama "Critique the architecture" --context src docs

Defaults and diagnostics

At initialization, PALABRE prefers nemotron-3-nano:4b if installed, otherwise the first installed model; without an installed model, that same name is the fallback. gemma4:e4b in the example is an explicit choice, not a mandatory default.

--set-ollama-model modifies only ollama-local and requires an installed model. For another agent, edit its model field and reapprove the file after reading it. palabre doctor checks server and model availability; model-unavailable signals a missing model, while network failure may have an HTTP diagnostic.

Advanced options: keepAlive, pullTimeoutMs (30 minutes by default) and maxOutputBytes. See Configuration.