Skip to content

Getting Started Overview

Murmur is the infrastructure layer for AI agents. It packages any agent as a single capsule — inference, tools, skills, and configuration all declared in one manifest — and runs it on your own infrastructure, fully isolated, with nothing calling out except what you declared.

What a capsule declares

A capsule manifest (murmur.yaml) is the single source of truth for an agent. It declares:

  • Which artifacts (tools, drivers, bootstrap logic) the capsule depends on
  • What capabilities are granted (filesystem, network, shell)
  • How inference is configured (transport, model, turns)
  • What outputs are expected

Nothing runs that isn't declared.

Why Murmur

Today's agent harnesses shift under you without warning, and every vendor ships its own opinions you have to work around. Murmur places the opposite bet: one manifest as the whole contract, defined by you for any occasion. Batteries included:

  • Isolated, WASM-sandboxed capsules
  • Secure by default, capability-scoped
  • Reproducible, pinned artifacts
  • Model-agnostic
  • Language-agnostic (WASM or native)
  • Composable from versioned artifacts
  • OCI-native (deploy to a VM or Kubernetes)

Design philosophy

Murmur is opinionated about interfaces and boundaries, not implementation details.

  • The runtime defines standard contracts (manifest shape, capabilities, interfaces)
  • Artifacts implement behavior (tooling, provider drivers, bootstrap logic)
  • Teams can use defaults first, then swap artifacts without rewriting the runtime

See Quickstart for the declare → install → run → inspect loop, and Language Guides for language-specific workflows (for example, building tools in Python).