Simulate the World.
Own the Infrastructure.
Murmura is the first open framework for enterprise swarm-intelligence simulation. Run millions of agents, survive any crash, connect to 400+ live data sources — and deploy entirely inside your own infrastructure.
import { bootstrap, resolveImport } from "@murmura/core";
import { SimulationQueues } from "@murmura/core";
await bootstrap({
triggersPath: resolveImport("./triggers"),
flowsPath: resolveImport("./flows"),
nodesPath: resolveImport("./nodes"),
queues: SimulationQueues,
});
Existing tools weren't
built for production
Every serious simulation team hits the same three walls. Murmura was built to eliminate all three.
Simulations crash. Costs explode.
A six-hour run with 10,000 agents dies at the final tick. Your team restarts from zero — re-invoking every LLM call, re-paying every API cost. Without durable execution, every failure is a full rebill.
Your data can't leave the building.
Government classifications, financial data residency rules, and defence security mandates rule out every cloud-hosted simulation tool. Procurement kills the deal before you write a single line of code.
The plumbing takes months.
Before the first simulation tick, every team rebuilds the same wiring: Temporal workflows, retry logic, observability, data connectors. Murmura ships it all pre-built, tested, and production-ready.
Murmura solves all three — out of the box. See how it works →
Three primitives.
Every simulation.
Whether you're modelling policy impact, market dynamics, or epidemic spread — every Murmura simulation is composed from the same three building blocks. Composable, independently testable, production-hardened.
Trigger
Detects an event — a webhook, file upload, API call, or schedule. It derives a unique SimulationId and hands off to a durable Flow. The simulation starts here and can never be silently lost.
Flow
The durable orchestration layer. It sequences every step, awaits external signals, and checkpoints each stage. If the process dies, Temporal replays from the last checkpoint — not from the beginning.
Node
The unit of real work. An LLM call, a simulation tick, a graph write, an API fetch. Each node runs with automatic retries, enforced time limits, full logging, and support for deterministic replay.
Example · Policy Impact Simulator — 2 million synthetic citizens
Engineered for the deployments
most tools refuse
Government procurement. Financial data residency. Classified networks. Defence audit requirements. Murmura was designed from day one for environments where failure is not an option.
Durable execution, guaranteed
Built on Temporal. If the host machine dies mid-simulation, Murmura resumes from the last checkpoint — automatically, without human intervention, with zero data loss or cost duplication.
Air-gap and on-prem ready
Deploy the entire stack inside your Kubernetes cluster. No data crosses your perimeter. Passes security review, financial data residency requirements, and defence procurement without exceptions.
Immutable audit trail
Every node call is signed, timestamped, and replayable. Regulators, auditors, and oversight boards can inspect exactly what each agent decided, in what order, and with what source data.
Swap any component
LLM provider, memory layer, simulation engine — each is independently replaceable at deploy time. Run GPT-4 on cloud and a local Llama model in the classified environment. Same orchestration code, zero changes.
A complete tool ecosystem
Versioned npm packages for every layer — source connectors, GraphRAG, the simulation engine, report synthesis. Pull in only what you need. Each package is independently versioned, tested, and published.
Security from the foundation
PII scanning on all inbound records, SSO via Authentik, OpenTelemetry distributed tracing, and Prometheus/Loki/Grafana observability. Supply-chain audited, SOC2-compatible architecture out of the box.
Wherever human decisions
shape the outcome
Murmura runs wherever scale, accuracy, and discretion are non-negotiable.
Policy & economic modelling
Run a proposed tax change through 2 million synthetic citizens across 50 demographic cohorts — before it becomes legislation. See second and third-order effects in hours, not quarters.
- 10,000+ agent populations
- Live legislative data via connectors
- On-prem, air-gapped deployable
Risk & market stress testing
Simulate thousands of adversarial market conditions simultaneously. Test portfolio exposure against emergent agent behaviour, contagion dynamics, and black swan cascades — not just historical data.
- Multi-scenario parallel execution
- Temporal durability for 24h+ runs
- Full audit trail per tick
Wargaming & crisis simulation
Model adversarial decision trees and friendly-force responses in a fully auditable, air-gapped environment. Local LLMs, zero cloud dependency, complete chain of custody from input to output.
- Fully air-gapped deploy
- Local LLM support
- Cryptographic replay proof
Social & epidemiological modelling
Ground your agents in real-world data from 400+ connectors. Reproduce and replay any simulation exactly. Publish with a deterministic proof — every peer reviewer can re-run your world from scratch.
- 400+ live data connectors
- GraphRAG-grounded agents
- Deterministic Temporal replay
Nothing else does
all of this. We checked.
Every alternative was built for a different problem. Murmura is the only framework that combines swarm simulation, durable execution, enterprise deployment, and live data — in a single open package.
| Capability | Murmura | MiroFish | Temporal | LangGraph | Prefect |
|---|---|---|---|---|---|
| Swarm / multi-agent simulation | ~ | ||||
| Durable execution — crash & resume | ~ | ||||
| On-prem / air-gap deployment | |||||
| Framework, not an app | ~ | ||||
| Enterprise audit trail | ~ | ||||
| Live data connectors (400+ APIs) | |||||
| GraphRAG-grounded agents | ~ | ~ | |||
| PII scanning & safety pipeline | |||||
| Provider-agnostic LLM & memory | ~ | ||||
| Versioned, composable packages |
Join the organisations
building with Murmura
We're onboarding a small initial cohort of enterprise teams, government agencies, and research institutions. If you're building anything that requires accurate, resilient, large-scale simulation — we want to hear from you.
No marketing. No spam. Just a direct conversation when your cohort opens.