AGENT INTERFACE / MCP 2026-07-28

Robots need an
agent contract.

Embodied Arena now exposes HILO knowledge as deterministic, structured, read-only tools. Agents can reason over benchmark evidence without gaining a path around physical safety.

Read the manifest ↗Connect an agent
01 / CLOSED LOOP

One trace. Two control planes.

The reasoning plane may interpret, compare, and propose. The robot plane alone may authorize and execute.

REASONING PLANEMCP + agent

Reads protocols and evidence, resolves intent, and produces typed proposals with provenance.

CONTROL PLANEIndependent safety kernel

Checks limits, state, permissions, stop conditions, and hardware interlocks before dispatch.

02 / TOOL SURFACE

Small tools, explicit meaning.

Stable names, strict JSON Schema, structured output, read-only annotations, and deterministic ordering reduce ambiguity and improve caching.

01compare_systems

Compare two to four named robot models or edge platforms while preserving layer boundaries.

READ ONLY · IDEMPOTENT
02get_hilo_protocol

Read HILO’s loop, Human Burden, MTHI, event vocabulary, safety split, and time tiers.

READ ONLY · IDEMPOTENT
03get_site_map

Choose the narrowest human or machine-readable source for the next agent step.

READ ONLY · IDEMPOTENT
04list_edge_platforms

Inspect Jetson-class compute as hardware—not as a robot policy or embodiment.

READ ONLY · IDEMPOTENT
05list_robot_models

Filter the public-beta seed set by task and open availability, with caveats attached.

READ ONLY · IDEMPOTENT
03 / BENCHMARK CONTRACT

Agent-native does not mean agent-controlled.

Every output carries the limits needed to stop an agent from promoting a convenient number into an unsupported claim.

01Provenance travels

Source, evidence type, protocol state, and public-beta caveats stay adjacent to the data.

02Layers stay separate

Model, runtime, edge compute, embodiment, and safety envelope are never silently collapsed.

03Human burden counts

Agents can retrieve Human Burden and MTHI definitions, not hide operator labor behind task success.

04Action requires authority

This public MCP server exposes no mutating or actuation tool. A future action server must be separately authenticated and approved.

04 / CONNECT

One endpoint, two examples.

Use the protocol request with any compliant client. The OpenAI Responses API is shown as one reference consumer, not a required runtime.

MCP / VENDOR NEUTRALSPEC ↗
POST /mcp
MCP-Protocol-Version: 2026-07-28
Mcp-Method: tools/call
Mcp-Name: get_hilo_protocol

{
  "jsonrpc": "2.0",
  "id": "hilo-1",
  "method": "tools/call",
  "params": {
    "name": "get_hilo_protocol",
    "arguments": {},
    "_meta": {
      "io.modelcontextprotocol/protocolVersion": "2026-07-28",
      "io.modelcontextprotocol/clientInfo": {
        "name": "my-agent",
        "version": "1.0.0"
      },
      "io.modelcontextprotocol/clientCapabilities": {}
    }
  }
}
OPENAI / REFERENCE CONSUMERDOCS ↗
const response = await client.responses.create({
  model: "gpt-5",
  tools: [{
    type: "mcp",
    server_label: "embodied_arena",
    server_url: "https://YOUR-DOMAIN/mcp",
    require_approval: "never"
  }],
  input: "Compare GR00T N1.6 with Jetson AGX Thor."
});

Because every exposed tool is annotated read-only, the OpenAI example disables per-call approval. Keep approval enabled for any future tool that writes data, contacts people, changes a benchmark submission, or proposes physical execution.

05 / EVOLUTION LOOP

Change deliberately.

A daily research review watches protocol, model, edge, and benchmark shifts. Publication still requires evidence and a human decision.

01WATCH OFFICIAL SOURCES02DIFF CAPABILITIES03ASSESS SAFETY + COMPATIBILITY04PROPOSE SITE CHANGE05HUMAN REVIEW06VERSION + PUBLISH
llms.txtmodel-readable guide ↗agent.jsonendpoint manifest ↗openapi.jsonbenchmark API ↗spec.jsonprotocol contract ↗