Docs

Local Models

Agenties can discover and use local Ollama models for supported agent and lab workflows. Claude Code remains the default execution path; local models are opt-in, reversible, and best treated as an experimental local backend until you validate a model against your own tasks.

Warning:Local models are not a drop-in replacement for Claude Code. Smaller local models may have limited context, weaker tool discipline, and slower first-run latency while Ollama loads the model.

What the local backend does

AreaBehavior
ProviderOllama is the supported provider today. Agenties talks to the configured Ollama base URL, usually http://localhost:11434.
DiscoveryAgenties can list installed Ollama models and surface them in model selectors as local options.
ExecutionSupported local tasks run on-device through the local model runner instead of the Claude Code CLI path.
LabsBuilt-in smoke, composer, code-tool, and coding-trial tests help determine whether a local model is usable before relying on it.
RollbackLocal task reports track touched files and rollback metadata for supported lab/task runs.

Setup

Open Settings, find the Local Models section, enable the local backend, and point Agenties at your Ollama server. Then refresh the model list or type a model name manually.

Ollama
ollama serve
ollama pull qwen2.5-coder:7b

# Agenties default base URL
http://localhost:11434

Health checks

Use the built-in health check before running a task. If Ollama is not reachable, Agenties keeps the local path unavailable and you can switch back to a Claude model from the same model selector.

Where local models appear

SurfaceWhat you can do
Chat model selectorPick a local Ollama model for supported local orchestrator/chat turns.
Team model selectorAssign a local model to a team member when local agent execution is suitable.
Agents tabRun a bounded local agent task and inspect the saved report.
Council previewLocal Ollama models can be offered as council participants when Council Mode is enabled.
MCP toolsUse localModels:* tools for health, listModels, labs, tasks, reports, and rollback.

Best practice

Treat local models as a capability you qualify per project. Run the labs first, check whether the model can follow file/tool constraints, and keep critical production changes on Claude Code unless the local model has passed the relevant coding trial.