Runtimes, Profiles, and Experts

MindMux separates discussion behavior from execution infrastructure.

Experts

An expert shapes how an in-app session behaves. Experts are part of the discussion layer. They define stance, reasoning style, or task framing within the product surface.

MindMux deliberately uses the term Expert in the chat layer to avoid conflating discussion roles with execution agents.

There are currently 7 built-in experts:

• general
• minimalist
• critic
• structured
• first-principles
• systems
• risk

Users can also create custom experts, stored in brain/experts.json.

Agents

Agent is the better term for execution contexts, especially dispatched work such as coding tasks or review tasks.

Runtimes

A runtime is the model execution system behind a session or task. Currently available runtimes:

• Claude Code: uses the anthropic protocol.
• Codex: uses the OpenAI Responses API.

Profiles

A profile is the model setup users choose in the app. It determines the runtime, model, and connection details needed for execution.

Built-in profiles reuse existing local login state, such as Claude Code CLI or Codex OAuth. Users can also add their own API-backed profiles in settings.

From a product-use perspective, the internal details do not usually matter: if a profile is shown as available, it can be used for sessions; when the task type supports that execution path, it can also be used for dispatched tasks.

Protocol boundaries

Protocol is a runtime property, not a free-form option on every profile:

• Claude Code uses the anthropic-compatible protocol.
• Codex uses the Responses API; Chat Completions is no longer supported.

Therefore, MindMux is not a generic "OpenAI-compatible ecosystem adapter". If a provider only offers Chat Completions, use that provider's anthropic-compatible interface and configure it as a Claude Code profile instead.

Why this separation exists

The knowledge layer should survive changes in model vendors and execution stacks. By separating brain from runtime, MindMux can keep the project memory stable while execution tooling evolves.

Model and effort selection

The runtime system also exposes model choice and, where supported, effort or reasoning-level controls. These are session-relevant settings rather than hidden implementation details.

Configuration as a first-class workflow

MindMux treats runtime and connector configuration as product-level behavior. The app includes a configuration-oriented expert flow so setup is not reduced to a raw settings form with no guidance.