Mantra OS — Mantra SDK

Multi-agent systems, built for production.

The framework behind every CloudMantra product: orchestrate crews of LLM agents with typed handoffs, graph memory, sandboxed execution, and a compliance layer that survives enterprise security review.

Talk to an Engineer
margin_watch.py
from mantra import Crew, Agent, Policy

crew = Crew(
    agents=[
        Agent("analyst", model="claude", tools=[erp.read]),
        Agent("executor", model="claude", tools=[erp.write],
              sandbox=True),
    ],
    memory="graph",  # facts · experiences · beliefs
    policy=Policy.preset("PCI-DSS")
         + Policy.human_escalation("finance-team"),
)

result = crew.run(
    "Flag margin leaks in Q3 invoices and draft corrections"
)
Python-first
Core framework and reference implementation
4 client SDKs
TypeScript, Go, Java, and .NET
Rust core
Optional high-performance execution engine
FastAPI SSE
Real-time streaming server, batteries included
Core Capabilities

Six things most agent frameworks skip.

01
Multi-agent orchestration

Compose crews of specialized agents with pluggable context-sharing strategies for how they divide, share, and hand off work.

02
Typed agent-to-agent handoffs

Schema-validated handoff protocols between agents — no free-text telephone games. Every handoff is structured, validated, and auditable.

03
Graph-based memory

Agents remember as a graph of facts, experiences, summaries, and beliefs — so systems get smarter about your business instead of starting cold every session.

04
Sandboxed execution

LLM-generated code runs in isolated sandboxes, never directly against your systems. Agents act through governed tools with full information-flow auditing.

05
Safety & compliance layer

A policy engine, human-escalation supervisor, and data-loss-prevention analyzers — with GDPR, HIPAA, and PCI-DSS presets and data-lineage tracking built in, not bolted on.

06
Model-agnostic routing

Claude, GPT, and local models through OpenAI, Anthropic, and Azure OpenAI endpoints — swap or mix models per agent without touching orchestration code.

Architecture

How a task flows through Mantra.

Every step is validated, policied, and audited before anything touches your systems.

Agent crews
Specialized agents, one goal
Typed handoffs
Schema-validated A2A protocol
Graph memory
Facts · experiences · beliefs
Policy engine
Compliance presets · escalation
Governed tools
Sandboxed writes to your systems
Task in → Validated actions outAudited end-to-end
Security

Guardrails that hold up in regulated environments.

LLM security proxy

A transparent proxy intercepts every model call: pre-request prompt-injection detection, post-response data-leak scanning, active PII redaction, and output validation. Detection is trained against the OWASP LLM Top 10.

Patent-pending secure architecture

A privacy-preserving communication architecture for distributed multi-agent systems handling sensitive data — designed for the security reviews enterprise and defense deployments actually face.

Compliance as presets

GDPR, HIPAA, and PCI-DSS ship as policy presets you turn on, not consulting projects. Information-flow auditing and data-lineage tracking come with them.

Everything in the compliance layer
GDPR presetHIPAA presetPCI-DSS presetHuman-escalation supervisorDLP analyzersPII redactionInformation-flow auditingData-lineage trackingRole-based accessFull audit trail

Ready to add a brain to your ERP?

A live 30-minute demo. No slides — your data, your decisions, your ROI.

Book a Demo