Contextual AI Data Movement
When an AI agent calls a real tool, data moves across multiple boundaries — model, gateway, API, and back. Aarmos makes that movement visible and contextual: you see what is leaving, where it is going, and why, before it happens.
Data flow, not data mystery
Most agent platforms treat data movement as plumbing. Aarmos treats it as a first-class concern — because the operator, not the vendor, should decide what crosses each boundary.
- Prompts go directly to your LLM. Chat traffic travels from your browser straight to the provider you chose — OpenAI, Anthropic, Google, Groq, OpenRouter, or a local Ollama runtime. Aarmos does not proxy, log, or inspect prompts server-side.
- Tool calls are intercepted and surfaced. When the model requests a tool, the gateway shows you the exact payload — arguments, destination, and capability class — before any network call is made. You decide; the agent waits.
- Credentials stay on the device. LLM API keys and OAuth tokens live in an AES-GCM vault in browser storage, wrapped with a KEK derived from your device password via PBKDF2-SHA256. The vault is unlocked only when you are actively using the app.
- Audit stays local. Consent decisions and run traces are written to an on-device ledger and signed as portable Receipts. The primary copy is always local — not a SaaS database you do not control.
Where every byte goes
These are the principal data surfaces. Aarmos has no server-side session store; opt-in product analytics and the narrow CORS proxies documented under Privacy notes are the only additional server touch-points.
Contextual, not just connected
Aarmos adds context to every data movement by labelling it with:
- Capability class — one of the 8 verbs (read, write, delete, execute, communicate, delegate, transact, control) — so you know the risk level at a glance.
- Initiating tool and agent — so you know which part of your workflow triggered the call.
- Delegation chain — when sub-agents are involved, the full hand-off path is recorded.
- Policy context — which rules matched, whether a signed org bundle was active, and what the operator decided.
That context travels with the receipt, so an auditor reviewing a JSONL export six months later can reconstruct not just what happened, but why it was allowed.
Aarmos governs the gateway. Once data leaves for your LLM provider or a connected MCP server, it is subject to their terms and security posture. Choose providers whose data handling you accept. For maximum control, run a local model via Ollama — then prompts and responses stay on the device.