Run · For the Platform Engineer — “Will it work?”
Run your existing AI stack — safely.
You already picked frameworks, models, and protocols. A Governance Runtime that forces a rewrite is one that never ships.
Aarmos sits inline as a microsecond-scale policy gate in front of every tool call — MCP, OpenAPI, deep-links, LangChain, CrewAI, AutoGen — without touching the agent code.
Decisions execute through tools. The gate is where a named decision, its participants, and its verdict meet the call that actually happens — which is why the receipt can be written at the moment of execution rather than reconstructed afterwards.
Read the Field Guide chapter →- BYO-LLM
- µs-scale p99 gate
- Signed receipts
- Offline verifiable
Protocols & frameworks
Two runtime modes
One runtime, two modes — both produce signed receipts, neither weakens a security boundary.
Transparent Runtime
Runs any framework, any HTTPS call, unmodified. Hostname-scoped receipts + policy enforcement. Aarmos does not terminate or decrypt TLS — it sees hostnames, not payloads.
Native Runtime Integrations
First-class protocols — MCP, OpenAPI, Browser, CLI, Deep-links — produce full execution evidence per tool call, chained into the signed receipt.
Framework integrations (LangChain, CrewAI, LlamaIndex, others) are community-owned and run unmodified through the Transparent Runtime.
How it works
- How the policy gate works →Governed tool calls pass through an inline gate. Predictable microsecond-scale p99, deny-override, tenant-scoped, atomic metering.
- CLI bridge — the arms →Scoped, deny-by-default access to files, shell, git, and read-only Postgres on the developer's machine.
- Local-first execution →Credentials, secrets, and receipts stay on-device. BYO-LLM (Ollama, Azure, OpenAI, Anthropic).