DocumentationAI agent runtime

MARS

MARS is the Multinex open-source runtime framework for agents that need identity checks, memory-aware planning, reviewable execution, and distributed handoff without locking the core runtime behind a hosted product.

What MARS provides

MARS gives builders a runtime structure for agents that need more than a prompt loop. It combines lifecycle stages, identity checks, memory-aware handoff, and a kernel surface that can be inspected and extended.

The open-source boundary is intentional: the framework is public, while hosted Multinex products, private connectors, customer data, and commercial policy packs stay outside the repo.

Runtime lifecycle

The core lifecycle is PERCEIVE, PLAN, EXECUTE, REFLECT. That shape makes each agent step easier to test, interrupt, approve, and hand off to another process or person.

  • Perceive current task, identity, context, and memory hints
  • Plan a bounded action path
  • Execute the smallest approved action set
  • Reflect outcomes back into journal or memory surfaces

Use MARS with MemQ and Legion

MARS is the runtime foundation. MemQ provides governed memory for continuity, and Legion can use MARS concepts when agent teams need commerce workflows, approvals, and operator packets.

Open-source quick start

Clone the public MARS repository, install Deno 2.x, cache the kernel, and run the SDK/runtime tests. Keep secrets out of examples and use local mocks for gateway behavior.

Setup
git clone https://github.com/multinex-ai/mars.git
cd mars/kernel
deno cache archon.ts
deno test --allow-read --allow-write --allow-net

How to evaluate MARS

Use this page to decide whether the product solves the immediate bottleneck: forgotten AI context, sensitive-data risk, or uncoordinated agent work. Then follow the lowest-risk setup path and prove one workflow before expanding.

Buyer checklist

MARS decision path

MARS is a fit when the value is clear to the user who has to do the work, not only to the team buying the platform.

  • Confirm the user problem this product removes first.
  • Run one practical setup path and verify the outcome.
  • Add the next Multinex layer only after the first proof is useful.