PYTHON · NUMPY

MCP for robots, done right

Expose a robot to a language model over a tool that enforces the safety envelope in the protocol.

01Challenge

Give an LLM a run_skill tool. Make the tool itself gate capabilities and enforce the envelope, so that a hijacked model gets a bounded trajectory and a safety report — and an incompatible robot gets a refusal.

02Model

ros-mcp proved you can hand a robot to an LLM over MCP. But a thin passthrough forwards whatever the model says. The Robot MCP Profile puts the safety envelope IN the protocol: run_skill binds through the runtime and returns a safety_report proving enforcement. The model is never trusted with unbounded actuation — the concrete answer to a hijacked agent driving a real robot.

03Guided practice

1. The tool. run_skill(name, robot, policy). 2. Gate first. Incompatible → refuse with reasons, no motion. 3. Enforce. Run the policy through the envelope. 4. Report. Return safety_report{within_cap, all_in_range, max_step} — the proof the server bounded it.

04Feedback

PASS WHEN PASS when run_skill bounds a hijacked policy and refuses an incompatible robot.

If your run fails, check:

Read the bench’s printed result — it passes when run_skill bounds a hijacked policy and refuses an incompatible robot. Run the cells top to bottom; if one errors, the FAIL line names the step to fix.
Gate before running; return a safety_report; refuse the mobile base.

05Retrieve & space

What's wrong with a thin MCP passthrough for robots? (it forwards whatever the model says — no safety).
What does run_skill return besides a trajectory? (a safety_report proving the envelope was enforced).
Why put the envelope in the protocol? (so a compromised model can never command unbounded motion).

06Mastery & project

You can expose a robot to an LLM through a tool that is safe by construction.

The habit: the protocol boundary is a safety boundary. Enforce there.

GO FURTHER · THE LIVE PLAYGROUND

The Robot MCP Profile + reference server ↗ — The same run_skill, as a real JSON-RPC MCP server.

← Compose LeRobot — one adapter, any checkpoint Capstone — author and ship a skill →

PYTHON · NUMPY

MCP for robots, done right

Build run_skill so the protocol itself is safe.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you