Agent Runtime

The Agent Runtime
for Vibe Coders.

Build fast. Run safe. Scale with control.

Agent
Prime
Engine
Mesh
Result

You built it in a weekend.

Now try running it in production.

No Structure

47 files changed. No record of what ran, where it ran, or why it ran there.

No Memory

Session ends. Context gone. Your agent starts over from zero every time.

No Gate

Everything ships. Nothing gets reviewed. Drift happens silently.

THE RUNTIME

One runtime. Three guarantees.

Agent Runtime

Structure every agent interaction. Dispatch, execute, return.

Structured Memory

CRDT-based blocks. Persistent across sessions. Conflict-free.

Gate Enforcement

GO / FIX / BLOCK. Every action gets a verdict before it ships.

From zero to runtime in four commands.

terminal
$ musu init
Initializing MUSU runtime... done.
$ musu run --agent claude
Agent connected. Prime orchestrating.
$ musu status
Prime: active | Engine: 3 workers | Mesh: 2 peers | Memory: 847 blocks
$ musu gate --mode strict
Gate enforcement: GO/FIX/BLOCK enabled.

Works with Claude Code, Cursor, Codex, and any MCP-compatible agent.

Five layers. One system.

Built for distributed AI execution.

PrimeOrchestrates
EngineExecutes
MeshDistributes
ControlGates
MemoryPersists

Prime

Central dispatcher. Capability-based routing. OODA loop.

Engine

BitNet inference. Worker pools. Job lifecycle tracking.

Mesh

P2P direct. QUIC + TLS 1.3. NAT traversal. No relay.

Control

GO / FIX / BLOCK verdicts. Intent zones. Audit trail.

Memory

CRDT blocks. Immutable runs. Conflict-free replication.

Not another AI framework.

Not a wrapper around LLM APIs.

MUSU operates below the model layer. It controls execution, not generation.

Not a prompt engineering toolkit.

MUSU enforces gates and structure. It doesn't help you write prompts.

Not a cloud service.

MUSU runs where you run. Your machine. Your network. Your data.

MUSU is infrastructure. It runs where you run.

Deploy MUSU.

Open source. Self-hosted. Yours.

In Active Development