Scoped memory
Tenant, workspace, user, role, and session lanes backed by indexed retrieval and leakage tests.
agent harness · governed execution
An agent harness for AI systems that have to keep running.
Muster wraps model calls, tools, memory, channels, MCP servers, and evals in one governed run loop so agents do not leak context, hide token spend, or learn without tests.
Controls
What the harness owns.
Token ledger
Every run records usage and flags replay waste so cost does not disappear into provider logs.
Tool policy
Capability packs and MCP servers declare permissions, secrets, result caps, and setup requirements.
Channel gateway
Terminal, web, Telegram, Slack, Google Chat, Teams, Discord, and WhatsApp surfaces enter the same envelope.
Eval-gated learning
Feedback becomes fixtures before behavior is promoted into durable agent behavior.
Where it fits
Use Muster around your agent framework, not only instead of one.
Muster is useful when your team already has prompts, tools, MCP servers, business apps, or model routes and needs governance around the run: memory boundaries, token visibility, setup checks, and reproducible eval evidence.