AI Settings

Configure AI models, agent behaviors, local AI, MCP server, and project memory.

Manage every aspect of how Omnilib's AI works — which models appear in the chat, how the agent behaves, what it remembers, and how it connects to external providers.

Access these settings at Settings > AI.

Model selection

The model selector shows all available AI models organized into three sections:

  • Cloud — Hosted models from Anthropic, OpenAI, Google, and others, routed through OpenRouter or direct APIs
  • Copilot — Models available through your connected GitHub Copilot subscription
  • Local — Models running on your machine via Ollama

Each model has a Show in chat toggle. Disable it to hide a model from the chat selector without removing it from your configuration. This keeps the chat dropdown uncluttered while preserving your settings.

Agent behaviors

Agent behaviors are custom instruction sets that change how the AI responds in specific contexts. For example, you might create a "Research assistant" behavior that emphasizes citations and structured summaries, and a "Code reviewer" behavior that focuses on brevity and correctness.

  • Create a behavior by clicking + New behavior, giving it a name, and writing the instructions.
  • Edit an existing behavior by clicking its name.
  • Delete a behavior from the behavior's edit view.
  • Use the AI Help button inside the behavior editor to get suggestions for improving your instructions.

Select the active behavior from the chat toolbar before starting a conversation.

Agent behavior mode

Toggle between Agent mode and Plan mode:

  • Agent — The AI executes tool calls directly as it works through your request.
  • Plan — The AI first outlines its plan and waits for your confirmation before executing any tools.

Plan mode is useful when you want to review the approach before the AI starts making changes.

Efficient Tool Execution

Toggle Efficient Tool Execution to enable MCP-based code execution mode. In this mode, tool calls are batched and executed more efficiently, reducing round-trips for multi-step tasks. This is recommended for complex coding and research workflows.

Project Memory

The Project Memory editor displays what the AI knows about your current project. Memory is organized into named sections (for example, "Architecture", "Conventions", "Open questions").

  • Switch between the structured form view and raw Markdown view using the tabs at the top.
  • Edit any section directly. Changes are saved automatically.
  • The AI reads project memory at the start of every conversation to provide context-aware responses.

AI Providers

GitHub Copilot

Connect your GitHub Copilot subscription to access its models in Omnilib.

  1. Click Connect GitHub Copilot.
  2. Omnilib starts an OAuth device flow and shows you a code.
  3. Open the displayed URL in your browser and enter the code.
  4. Authorize the connection.

Once connected, Copilot models appear in the Copilot section of the model selector. Click Disconnect to remove the connection.

Local AI (Ollama)

Run models locally on your machine using Ollama.

Omnilib displays a hardware summary showing your CPU, RAM, and GPU to help you choose models your machine can run comfortably. Based on this summary, Omnilib suggests compatible models.

  • Click Download next to any model to pull it via Ollama.
  • Click Delete to remove a downloaded model and free disk space.
  • Enter a model name in the Custom model field and click Add to use any Ollama-compatible model not listed by default.

Desktop only. Local AI requires Ollama installed on your machine. Download it from ollama.com.

MCP Server

Omnilib can expose an MCP (Model Context Protocol) server that lets external AI clients connect to your project's tools.

  • Enable/disable the MCP server with the toggle.
  • Set the port the server listens on (default: 3282).
  • Follow the setup instructions shown for each supported client:
    • Claude Desktop
    • Cursor
    • VS Code
    • Zed
    • Windsurf
    • Continue
    • Cline

Once connected, the external client can use Omnilib's agent tools through the MCP protocol.

Related