For Platform Engineering
Written for the platform engineer

For Platform Engineering.

Aarmos is a runtime you add next to your agents, not a platform you migrate to. No rewrite, no new framework. One agent, one policy, one Receipt — then scale from there.

1

Integration surface.

Aarmos sits at the action boundary. Your agent keeps its framework. You point it at the local runtime and get policy decisions plus signed Receipts as part of the local runtime — no separate service call, no metered API round-trip.

  • CLI: install, init, run — nothing runs in the background unless you start it.
  • Frameworks: LangGraph, CrewAI, raw HTTP, custom.
  • Protocols: MCP, A2A, OpenAPI, deep-links, custom APIs.
  • BYO-LLM: Ollama, OpenAI, Anthropic, Azure.
Read the quickstart
2

Rollout pattern.

Start with one agent. Write one policy. Watch one Receipt. Then a second agent. Then a third. Adopt it the way you adopted feature flags — one surface at a time, reversibly.

  1. 1Pick one agent — ideally one that touches an external tool.
  2. 2Write one policy — allow the safe path, deny the dangerous one.
  3. 3Run it — every decision now has a Receipt.
  4. 4Replay a Receipt — confirm the decision is reproducible.
  5. 5Add the next agent when your team is ready.
3

Day-two operations.

Aarmos runs on the machine the agent runs on. Health is observable locally. Incidents are answered by replaying the Receipt that matters, not by combing through logs.

  • Kill switch — stop an agent surface without a deploy.
  • Replay — rerun any decision against the pinned policy.
  • Receipts — the primary evidence for post-mortems.
  • Live view — `aarmos observe` streams a run to your terminal or browser over loopback, read-only, joined with a pairing code.
  • No shared control plane — no cross-tenant blast radius.

Then: understand, then improve

Every brokered call leaves a signed Receipt. Operations reads that stream to show a flight recorder of recent work, trust posture across installs, and counterfactual replay — so platform owners can test a policy change against real history before rolling it out.

Explore Operations

Ready to try it?

The pilot page walks the rollout in about thirty minutes, on your machine.

Run the pilotRead the quickstart
GovernCan this agent do this?RunWill this work with my stack?ProveCan I prove what happened?
One decision. Three checkpoints. Every time.