Orchestrate teams of AI agents
Marsys lets you design and run teams of specialized AI agents that work together on complex workflows—so you can move from isolated experiments to reliable, end-to-end automation.
Customer Support System
From One-Off Assistants to AI Teams
An individual "AI assistant" is great for answering questions, but most high-value work looks more like a team than a single role. Today, many organizations stitch this together with long prompt chains and manual review.
Multiple specialized agents
Each agent plays a clear role—research, analysis, checking, reporting—instead of one generalist trying to do everything.
Structured collaboration
See exactly who does what and when. No more black-box prompt chains or unclear handoffs between steps.
Built-in oversight
Humans can review decisions where it matters. Control and governance are baked into the workflow design.
Pragmatic Infrastructure for AI Teams
Marsys is an open-source framework that focuses on three things:
Designing workflows you can understand
Represent complex processes as topologies of agents—pipelines, hub-and-spoke, meshes—instead of opaque chains of prompts. This makes reviews, audits, and changes much easier.
Running many agents in parallel safely
Coordinate parallel branches with clear rules and state, so you can reduce cycle times without losing control of what happens where.
Keeping options open
Use your preferred models and data stack—OpenAI, Anthropic, Google, Groq, local models, LlamaIndex or other RAG systems—while Marsys focuses on orchestration.
Built for Builders, Ready for Scale
Whether you're writing the code or designing the strategy, Marsys bridges the gap between prototype and production.
For Developers
Build serious multi-agent systems without reinventing the runtime.
With Marsys, you can:
- Turn ideas like 'a research team with a reviewer and reporter' into concrete, reusable workflows
- Avoid writing yet another orchestration layer—no more custom queues, state machines, and ad-hoc retries
- Keep agent code clean and testable, while Marsys handles memory, tools, and coordination
- Start from simple patterns and grow into patterns as your use cases evolve
For Businesses
Put AI teams to work on your hardest workflows.
Marsys helps you:
- Move beyond 'AI demos' and pilot AI teams on real processes like claims review or research
- Shorten cycle times by running many steps in parallel, while still keeping humans in control
- Keep your data and models where you need them (cloud or on-prem) with open-source infrastructure
- Start small: experiment with one or two AI team workflows, learn from the results, and expand
Built for Complex, High-Stakes Workflows
Marsys excels where single agents fail—coordinating specialized expertise across regulated industries that demand auditability and control.
Finance & Insurance
Transform how financial institutions process complex workflows—from claims automation to fraud detection and compliance reviews.
Example Workflows
- Claims processing with parallel fraud detection, valuation, and policy checks
- Loan underwriting with multi-agent risk assessment and document verification
- Regulatory compliance reviews across multiple jurisdictions
- Portfolio analysis with concurrent market research and risk modeling
Intellectual Property & Patents
Accelerate patent research, prior art searches, and IP portfolio management with teams of specialized AI agents.
Example Workflows
- Prior art searches across multiple databases and jurisdictions in parallel
- Patent landscape analysis with concurrent technology and competitor mapping
- Freedom-to-operate assessments with coordinated legal and technical review
- IP portfolio strategy with automated classification and gap analysis
Legal & Regulatory
Handle complex legal analysis, document review, and due diligence with AI teams that coordinate specialized expertise.
Example Workflows
- Contract review with parallel clause analysis, risk assessment, and compliance checks
- Due diligence for M&A with concurrent financial, legal, and operational analysis
- Regulatory filing preparation with coordinated document generation and validation
- Legal research with multi-jurisdictional case law and statute analysis
Production-Ready Orchestration from Day One
Enterprise-grade features built into the open-source framework. No infrastructure work required.
Dynamic Branching & Parallelism
50+ agents in-flight simultaneously with adaptive branching. No artificial limits on concurrent execution.
8 Topology Patterns
Hub-and-spoke, pipeline, scatter-gather, and more. Pre-built patterns for common orchestration scenarios.
Multi-Provider Support
Works with OpenAI, Anthropic, Google, Apertus, and local models. No vendor lock-in.
State Management
Built-in checkpoints, pause/resume, and error recovery. Never lose progress on long-running workflows.
Data Sovereignty
Deploy on-premises or any cloud. Swiss/EU model support for FINMA, HIPAA, and GDPR compliance.
Multilingual Support
Native support for German, French, Italian, Romansh, Swiss German, and 1,000+ languages via Apertus.
Focused on Orchestration, Built for Production
Marsys solves the "how do we coordinate many agents over complex workflows" problem—and stays compatible with your existing LLM providers and RAG stack.
Marsys Unique Capabilities
| Capability | Marsys | Others(LangGraph, CrewAI, LlamaIndex) |
|---|---|---|
Branch Isolation True parallelism with branch-scoped state and memory | ||
Pure Agent Logic Agents implement pure _run(), orchestration handled separately | ||
Topology as First-Class 3-way definition (string, object, patterns) with 8 pre-built patterns | ||
Convergence Policies Built-in policies for parallel branch reconciliation | ||
Rules Engine Policy-based execution guards and resource limits | ||
State Persistence Snapshot/restore workflows with pluggable backends |
When Marsys is the Right Fit
Marsys is Designed to Coexist
Rather than replacing your entire stack, Marsys focuses on the orchestration layer. Use LlamaIndex or another RAG stack for retrieval, leverage LangChain integrations for tools and adapters, and choose your preferred models and cloud platform.
Marsys coordinates your agents—you control everything else.
Built to Support a Broader Platform
The open-source framework is designed to be the engine for more than just code-based projects. Over time, we plan to layer on hosted services and visual tools.
Today
Open-Source Framework
Apache 2.0
The core open-source library that provides agents, topologies, orchestration, tools, rules, and state management—available now.
- 8 topology patterns out of the box
- 50+ parallel agents in-flight
- Multi-provider (OpenAI, Anthropic, Google, Apertus, local)
- Full control, zero lock-in
Future
Hosted Control Plane
Planned
A future hosted service could provide a managed control plane for the framework with multi-tenant execution and enterprise features.
- Managed topology storage & runs
- Real-time monitoring & analytics
- WebSocket/event interfaces
- Foundation for visual tools
Future
Visual Builder
Planned
A future visual tool could sit on top of a hosted control plane, aimed at product and domain experts who want to design AI teams without writing Python.
- Visual topology designer
- Template library
- Run dashboards
- Collaboration features
Because these future capabilities will be built on the same abstractions you use today—agents, topologies, rules, state—you won't have to rewrite your workflows to take advantage of them.
Start Building with AI Teams
The open-source framework is available now. Join developers and enterprises building reliable multi-agent workflows.
For Developers
Install the framework, explore the documentation, and start building your first AI team today.