v1.10.3 · 5 §22–26 endpoints live ≈ 10-minute walkthrough · one install

Wire any robot to Claude —
get a signed audit bundle.

One file per robot. The plugin wires it into Claude Code; the gateway turns Claude's plans into deterministic actions and emits a RCAN-signed bundle of every action it took. Safety is enforced at Layer 3 — declaration alone doesn't move motors. 

$ claude plugin add robot-md
# or, without a Claude Code MCP client
$ pipx install robot-md && robot-md init
1 robots registered
7 peer runtimes
5 §22–26 endpoints live
v1.4.0 on PyPI
§ 01 — The closed loop

ROBOT.md → MCP → Claude Code → gateway → robot moves.

The full loop — manifest declaration through Layer 3 enforcement to physical action — all in one conversation. The gateway is what verifies and moves; declaration alone never does.

~/bob · robot-md · MCP session
$claude plugin add robot-md
✓ robot-md plugin installed · MCP wired
$robot-md init
✓ ROBOT.md + CLAUDE.md written — preset so-arm101
✓ registered — RRN-000000000003
✓ calibrated — zero pose recorded via IK
$claude
✓ robot-md-mcp connected · 6 resources · 3 tools loaded
You:pick the red lego and place it in the bowl
Claude:Scanning workspace… found red 2×4 lego at (0.31, −0.12, 0.04)
→ IK solved · executing pick …
30-second screencast placeholder
bob arm · pick and place
Step 1
ROBOT.md
YAML frontmatter + prose body declares what the robot is and what it can safely do
Step 2
robot-md-mcp
MCP server exposes 6 resources + 3 tools to any connected agent
Step 3
Claude Code
Reads manifest, reasons over capabilities and safety gates, plans the action
Step 4
Robot moves
Driver executes; E-stop and HiTL gates enforced per §8 AUTHORIZE before servo bus
30-sec walkthrough →
§ 01a — Architecture

Where ROBOT.md fits

ROBOT.md is the declaration layer — one of six in the OpenCastor stack. It declares what a robot is, what it can do, and what limits hold. Declaration is not enforcement.

  • Layer 1 — Declaration: ROBOT.md, robot-md CLI, robot-md-mcp (Advisory).
  • Layer 2 — Agent runtime: Claude Code, Codex, Gemini, any MCP host.
  • Layer 3 — Gateway / Enforcement: robot-md-gateway. Mandatory exclusive path.
  • Layer 4 — Robot-facing runtime: OpenCastor and other RCAN runtimes.
  • Layer 5 — Protocol: RCAN spec + SDKs.
  • Layer 6 — Registry: Robot Registry Foundation.

Pattern 1 (Declaration-only): install robot-md alone for manifest authoring, schema validation, and dry-run agent planning. No physical robot is moved.

Pattern 2 (Gateway Mode): add robot-md-gateway when you actually want an agent to drive hardware. See live version compatibility →

§ 02 — Surface map

One file. Every surface.

First-class on Claude Code; works with any MCP-aware agent. Install once per surface, then ask in plain English.

● Live
Claude Code
claude plugin add robot-md
Full install guide →
● Live
Gemini CLI
gemini mcp add robot-md \
  npx -- --yes robot-md-mcp \
  ./ROBOT.md
Full install guide →
● Live
OpenAI Codex
codex mcp add robot-md \
  -- npx --yes \
  robot-md-mcp ./ROBOT.md
Full install guide →
○ Pre-release
ChatGPT
npx -y robot-md-http
# OpenAPI bridge for
# Custom GPT Actions
Full install guide →
◌ Pre-release
Amazon Q
# MCP: pending aarch64
# robot-md-http bridge
# available today
Full install guide →
§ 03 — Compliance moat

Signed evidence. Submitted to RRF. Not just generated.

Five RCAN §22–26 attestation packets, all live endpoints. Each packet is produced by a robot-md emit-* command, signed with the operator's release key, and submitted to RRF intake at /v2/<packet-type>. RRF stores the evidence in its transparency log; jurisdictional filing remains the operator's responsibility.

§22
FRIA
Fundamental Rights Impact Assessment — automated pre-deployment checklist filed to RRF.
Live
§23
Safety benchmark
ISO 42001 self-assessment linked to the robot's RRN. Signed envelope, audit-trail hash.
Live
§24
IFU Art. 13(3)
Instructions for Use — machine-readable intent and capability declaration per Article 13(3).
Live
§25
Incident report
Structured incident reporting per Article 72, linked to RRN, timestamped and signed.
Live
§26
EU Register Art. 49
Per-model EU AI Act registration packet, routed to the official register endpoint via RRF.
Live

All five endpoints live as of 2026-04-24. Read the full compliance story → RRF registry details →

§ 04 — Founder-authored protocol

robot-md is the reference declaration layer
for the RCAN protocol.

RCAN (Robot Communication & Addressing Network) is the open wire protocol for signed, authority-aware robot messages. robot-md's team authored the spec. That means every breaking decision — from ML-DSA-65 crypto to the §22-26 EU compliance hooks — was made in-house, not adapted from a third-party interface. When the spec evolves, robot-md ships the declaration-format updates first; the reference SDKs (rcan-py, rcan-ts) ship the protocol-implementation updates.

rcan.dev/spec/ — read the protocol spec →
§ 05 — Built on · plugs into

Open standards. Zero runtime lock-in.

AI surfaces
Anthropic

Claude, MCP (Model Context Protocol), Managed Agents. Claude Code is the primary tested surface; other Claude clients work where MCP is supported.

claude.ai →
Infrastructure
Cloudflare

The Robot Registry Foundation runs on Cloudflare Pages + Workers + D1. Edge-served globally, zero downtime at the identifier layer. No seat fees. Self-hosted mode for private fleets.

robotregistryfoundation.org →
Open standards
RCAN + MCP

RCAN protocol (Apache-2.0, community-governed) for the wire protocol. MCP for agent-surface interop. JSON Schema draft 2020-12 for frontmatter validation. No vendor-specific formats in the core.

rcan.dev →
01 — Wire it

Do the cookbook.

Install robot-md, author a ROBOT.md, register with RRF, run a skill, and walk away with a signed audit bundle. About ten minutes, end to end.

Open the cookbook →
02 — Read it

Understand the format.

The ROBOT.md spec defines every required and optional field, the RCAN protocol conformance rules, and the compliance packet filing requirements. Normative text, schema, and examples.

Read the spec →
03 — Want this managed?

Compliance-bot waitlist.

Give your robot repo to Compliance-bot and get back a signed FRIA + IFU + safety-benchmark + EU-register filing in one day. Powered by Anthropic Managed Agents.

Join the waitlist →
Where safety is actually enforced.
Physical safety is enforced at Layer 3 (robot-md-gateway) or Layer 4 (a runtime that embeds it, e.g., OpenCastor). Declaration alone (Layer 1) does not enforce safety. Agent host alone (Layer 2) is not the safety boundary. If a deployment lacks Layer 3, no safety claim attaches to it.