Protocols¶

Jarvis adopts an open protocol stack to guarantee interoperability with third-party tools, agents and UIs without vendor lock-in.

The three agent protocols¶

Model Context Protocol (MCP)¶

Open standard introduced by Anthropic to let LLMs access external tools and resources in a structured, safe way.

flowchart LR
    AGENT[Agent / LLM] -->|MCP request| SERVER[MCP Server]
    SERVER --> TOOL1[tool: file_search]
    SERVER --> TOOL2[tool: gh.create_pr]
    SERVER --> RES1[resource: repo files]
    SERVER --> RES2[resource: docs]

In Jarvis every specialised agent exposes its own MCP server. Examples:

• agents/medical-agent/mcp_server.py → tools oura.fetch_sleep, whoop.recovery, fhir.observation_query
• agents/scraping-agent/mcp_server.py → tools crawl.fetch, firecrawl.crawl, jina.read
• agents/maker-agent/mcp_server.py → tools printer.start, slicer.slice, blender.export_stl

Apple at WWDC 2025 integrated MCP as the protocol for Siri/App Intents.

Agent-to-Agent (A2A)¶

Standard introduced by Google and later moved under the Linux Foundation for neutrality. Specifies how heterogeneous agents can communicate and delegate tasks.

flowchart LR
    ORCH[Orchestrator Agent] -->|A2A delegate| MED[Medical Agent]
    ORCH -->|A2A delegate| FIN[Finance Agent]
    ORCH -->|A2A delegate| WEB[Scraping Agent]
    MED -->|A2A response| ORCH
    FIN -->|A2A response| ORCH
    WEB -->|A2A response| ORCH

LangGraph v1.0 supports A2A natively as transport. Google ADK and Pydantic AI ship A2A SDKs.

Agent-UI Protocol (AG-UI)¶

Lets agents project reactive UIs directly to the frontend, without pre-compiled pages.

In Jarvis: the agent can generate on-the-fly a biometric dashboard widget, a filtered transaction table, a chart — all via AG-UI.

Layered transport stack¶

┌──────────────────────────────────────────────────┐
│   Application protocols                          │
│   MCP · A2A · AG-UI                              │
├──────────────────────────────────────────────────┤
│   Transport                                      │
│   HTTPS · WebSocket · gRPC · MQTT                │
├──────────────────────────────────────────────────┤
│   Auth & identity                                │
│   OAuth 2.0 / OIDC · JWT · FIDO2 · passkey       │
├──────────────────────────────────────────────────┤
│   Smart-home transport                           │
│   Matter · Thread · Zigbee · BLE                 │
├──────────────────────────────────────────────────┤
│   Health interop                                 │
│   HL7 FHIR · SMART on FHIR · IEEE 11073          │
└──────────────────────────────────────────────────┘

Data schemas¶

Conversation turn¶

{
  "turn_id": "uuid",
  "user_id": "uuid",
  "device_id": "uuid",
  "timestamp": "2026-05-09T12:00:00Z",
  "message": "How long did I sleep yesterday?",
  "language": "en",
  "context": {
    "location": "home",
    "modality": "voice",
    "active_focus": null
  }
}

Device registration¶

{
  "device_id": "uuid",
  "owner_id": "user_id",
  "device_type": "watch",
  "model": "PineTime",
  "capabilities": ["heartrate", "notifications", "haptic", "wakeword"],
  "trust_level": "primary",
  "paired_at": "2026-05-09T12:00:00Z"
}

Memory record¶

{
  "id": "uuid",
  "user_id": "uuid",
  "scope": "user|agent|session|org",
  "type": "fact|preference|event",
  "text": "Birthday: March 15",
  "embedding": [0.012, ...],
  "metadata": {"source": "conversation", "ttl": null}
}

Federated identity¶

For multi-instance environments (a user with a home server + work server):

• OIDC Federation between instances
• WebAuthn as portable second factor
• Passkeys synced via iCloud Keychain / Google Password Manager

Reference specifications¶

• Model Context Protocol — modelcontextprotocol.io
• A2A — agent2agent.dev
• AG-UI — ag-ui.com
• Matter — csa-iot.org
• HL7 FHIR — hl7.org/fhir
• OAuth 2.0 — oauth.net/2/
• WebAuthn — w3.org/webauthn