Docs

Runtime and MCP Architecture

MindMux separates model execution from project knowledge through a runtime layer and an MCP tool layer.

Runtime abstraction

The runtime abstraction allows different execution backends to power the same product surface. Chat sessions and tasks both rely on this layer.

The main goals are:

  • runtime portability
  • shared product behavior across backends
  • clearer boundaries between product logic and vendor-specific execution

Currently Supported Runtimes

The product direction is to let multiple runtimes operate against one knowledge model. The currently documented available runtimes are:

  • Claude Code
  • Codex

The important architectural point is not the brand list itself. It is that the brain does not belong to any one of them.

MCP as the knowledge boundary

MindMux exposes structured project capabilities through MCP tools rather than relying only on raw prompt context.

Important tool families include:

  • brain read and write tools
  • root doc tools
  • skill tools
  • task tools
  • configuration tools

Typical request path

Chat scope vs task scope

The tool surface is intentionally scoped. Some tools are available to the main chat agent but not to delegated task execution, especially where path resolution, orchestration authority, or project-wide writes need tighter control.

Examples of chat-only behavior include:

  • project-wide orchestration
  • user-interaction cards
  • broad knowledge writes

Examples of task-scope behavior include:

  • focused read access
  • constrained review or execution flows

Independent MCP server process

The codebase uses an MCP server boundary as a separate process rather than folding everything into one monolithic runtime loop. This supports stronger control over tool exposure and keeps the knowledge interface explicit.

External connectors and bridging

MindMux is also an MCP client. That means the product can both:

  • expose its own capabilities to runtimes
  • consume external MCP servers as connectors

Where a runtime cannot directly consume a given transport shape, the architecture includes a bridge layer that normalizes the connector into a usable form.

Why this matters

This architecture gives MindMux a stable knowledge contract. Runtimes can vary, but the project brain and its operations remain legible and structured.