Explanation

How muster works and why it is built the way it is. These pages are background reading; the how-to guides tell you what to type.

Concepts

• The problem muster solves: why one endpoint and a discovery layer, rather than every MCP server in every agent.
• Architecture: the aggregator, the meta-tools, sessions, the service locator that holds the code together.
• MCP aggregation: registration, naming, families, tool visibility per session, execution.
• Service orchestration and workflows: the lifecycle of a registered server and how workflows execute.
• Observability: what muster contributes to a trace, the metrics it exports and the structure of its logs.
• Design principles: the rules the codebase follows.

Components

• Aggregator: registry, tool factory, event handler and the stores behind a session.
• Workflows: definitions, execution engine and templating.

Decisions

The architecture decision records capture the choices that shaped muster: the service locator, OAuth in front of muster and towards remote servers, session-scoped tool visibility, single sign-on by token forwarding, server-side meta-tools.

Protocol analysis

MCP 2026-07-28 is muster's section-by-section analysis of the protocol revision that removes the session handshake and makes extensions first class: what each change means for an aggregator, what has to change, and what is still open.