Command Palette

Search for a command to run...

UnylyUnyly
Browse all

Norirobotics

FreeNot checked

MCP server + webapp for the Nori Robotics A3 home robot (real and virtual) — teleop, LeRobot recording, a live 3D rig viewer, and VR spawn into Resonite, Overte

GitHubEmbed

About

MCP server + webapp for the Nori Robotics A3 home robot (real and virtual) — teleop, LeRobot recording, a live 3D rig viewer, and VR spawn into Resonite, Overte, Godot, Unity3D

README

Just Ruff Python FastMCP Three.js 3D viewer MIT

MCP server + webapp for the Nori Robotics A3 — a $1,688, 19-DOF wheeled bimanual home robot (real and virtual). Control it from Claude, watch a live 3D rig viewer of the A3 right in the webapp, and spawn its virtual twin into Resonite, Overte, Godot, Unity3D — using other fleet repos (robotics-mcp + VR bridges, not built into this repo alone).

What this wraps

Wraps nori-sdk (Apache-2.0), Nori Robotics' official Python client. The SDK talks to the robot over a WebRTC data channel with Supabase Realtime signaling — there is no serial/USB/ROS bridge to fake. See docs/WRAPPEE.md for what Nori A3 is, its XLeRobot/Hugging Face LeRobot lineage, and where its community lives.

No physical A3 exists in this household yet — the robot ships Fall 2026. Every session-gated tool defaults to nori_sdk's own upstream-supported mock_session() until real Supabase credentials are configured. This is the SDK's declared mock, not a fleet-invented fake — see docs/ONBOARDING.md.

What You Can Do

How it runs: headless FastMCP server (stdio or HTTP) that opens one WebRTC session against either a real A3 (Supabase credentials configured) or nori_sdk.mock.mock_session(). Nori Robotics' own lab.norirobotics.com webapp is never bundled or replaced — this dashboard is a narrower, agent-oriented view (session status, tool calls, recording state).

Direction Artifacts Notes
Hands-in joint/gripper targets, Cartesian poses, jog payloads, episode task labels Via MCP tool params or the webapp Control/Recording pages
Hands-out robot telemetry/status/camera_layout, LeRobot-format episode recordings, video snapshots nori_session, nori_recording
  • Specs, SDK links, XLeRobot/HF LeRobot lineage, and Hacker News launch-thread reaction — no session required (nori_info)
  • Open/close/inspect a live control session, and manage named robot profiles — Virtual Twin plus any number of registered physical A3s, each response explicit about which one produced it (nori_session)
  • Jog, move-to-target, Cartesian pose control, e-stop/fault reset, and external-policy/leader-arm streaming (nori_control)
  • Named waypoint navigation — remember/navigate to a saved destination, Nav2 goal lifecycle (nori_navigation, new in nori-sdk 1.1.0)
  • Opt-in LiDAR/IMU streams and vision-stack world-state reads (nori_perception, new in nori-sdk 1.1.0)
  • Start/stop episode recording in LeRobot-compatible format, video snapshot/bitrate control (nori_recording)
  • VR/physics twin spawning — Unity/Overte/Godot/MuJoCo/Isaac via other fleet repos (nori_vr)
  • Live 3D viewer of the A3 right in the webapp — the correctly-posed rig, rendered from a real glTF/GLB export (73,974 verts / 134,656 tris), with wireframe toggle, a wave-demo animation, and orbit lighting (web_sota/src/pages/ViewerPage.tsx, Three.js-based BotViewer)
  • Fleet-aware: registers with robotics-mcp, pairs naturally with teleoperator-mcp for VR-driven demonstration collection, and feeds vla-mcp's LeRobot training pipeline

Part of the sandraschi Robotics + VR Fleet

norirobotics-mcp is one member of a larger set of MCP servers under the sandraschi account that work together rather than in isolation — a physical robot integration is far more useful with a virtual-first testing path and a shared control hub than as a standalone wrapper. This is deliberately not a solo project; it's built to the same portmanteau-tool, live-verified, honest-mock-vs-real conventions the whole fleet follows, and it's designed to plug straight into it.

Robotics fleet peers (declared in nori_info(operation="fleet_peers") / GET /api/hero, and reciprocally, norirobotics-mcp is registered as a member robot inside robotics-mcp itself — see that repo's robotics.nori_a3 config block and NoriMcpClient HTTP bridge):

Repo Role
robotics-mcp Fleet hub for physical + virtual robots — one unified robot_control interface across Dreame vacuums, Yahboom ROSMASTER, Unitree quadrupeds, drones, and Nori A3, plus virtual robot ("vbot") spawning into the VR platforms below. norirobotics-mcp registers here as a member robot, bridged via HTTP rather than reimplemented — see robot_control(robot_id="nori_a3", action=...).
teleoperator-mcp WebXR teleop gateway (Pico 4 / Quest). Nori A3's own control path is WebRTC remote-teleop — a natural pairing for VR-driven demonstration collection: teleoperate in VR, record the episode through this server.
vla-mcp Vision-language-action training pipeline. Currently alpha/shelfware fleet-wide — Nori A3's LeRobot-format episode recordings (nori_recording) are its first plausible real workload rather than a synthetic one.
universal-actuator-mcp Motor/actuator abstraction layer — the eventual home for Feetech-to-QDD actuator-upgrade tooling, a real gap the Hacker News launch thread flagged for the A3's RC-servo arms (see nori_info(operation="actuator_upgrade") for the sourced, no-specific-recommendation-yet note).
bumi-mcp Closest structural precedent in the fleet: another wheeled consumer robot, specs+OSS-info tools shipped first, physical control gated behind a verified bridge second — the same honest, staged rollout this repo follows.

VR crossconnects — using other fleet repos (how a Nori A3 gets a virtual twin via robotics-mcp + VR bridges — not standalone; this repo supplies the mesh, the fleet does the spawning): robot_virtual/vbot_crud tools — platform="unity" etc.):

Repo What it adds
resonite-mcp Social VR platform control via ResoniteLink (real-protocol WebSocket, not OSC). Ships resonite_link_spawn_fixture for gripper/manipulation test fixtures (box/cup/ball/table/chair) and resonite_link_animate for spin/bob/real-physics-bounce — a natural staging ground for testing a Nori A3 pick-and-place task in VR before running it on hardware.
overte-mcp Open-source metaverse platform control, the source these fixture-spawner/animate/model-depot patterns were first built and live-verified against before being ported to the other VR repos.
unity3d-mcp Unity Editor automation via a live TCP bridge — robotics-mcp's primary virtual-robot spawn target, now with the same fixture-spawner/animate capability (built from Unity's own native primitives, not a custom mesh generator).
godot-mcp Godot 4 engine control via TCP bridge — same fixture-spawner/animate pattern, plus a real model/texture asset depot with backup/restore.
vrchat-mcp VRChat integration - OSC avatar control plus REST (friends, users, notifications, real-time Pipeline events). VRChat's platform doesn't allow live external world-authoring the way the others do, so this one plugs in at the social/telepresence layer rather than the object-spawning one.

The fixture-spawner/animate/depot capabilities listed above landed across all four world/engine platforms in the same session (2026-09), each adapted to that platform's actual architecture — same closed-form bounce physics ported four times, verified identical, not four independent guesses. See each repo's own CHANGELOG.md for the live-verification details.

How the A3 actually gets spawned into VR — using other fleet repos

This isn't a generic placeholder box standing in for the robot — it's the real A3 mesh, spawned via other fleet repos, not this repo alone. scripts/export_posed_mesh.py takes the same posed rig the webapp's 3D viewer renders and emits three artefacts: nori_a3_posed.glb (single flattened mesh — Unity/Overte/Godot import), nori_a3_rig.glb (per-body hierarchy), and nori_a3_posed.mesh.json (26,467-triangle decimated mesh-JSON for spawn_mesh). The fleet hub robotics-mcp robot_virtual/vbot_crud (robot_type="nori_a3", platform="resonite"|"overte"|"godot"|"unity") picks that mesh up and pushes it live:

  • Resonite — resonite-mcp ResoniteLink.spawn_mesh (real WebSocket, verified)
  • Overte — overte-mcp (verified live, you saw it)
  • Unity3D — unity3d-mcp TCP bridge import_model + spawn_fixture (now real — nori_a3_posed.glb via model depot, not a box primitive)
  • Godot — godot-mcp TCP bridge + model depot (same GLB)
  • MuJoCo — local mujoco viewer (mujoco.viewer.launch_passive) from the same URDF — no fleet repo needed, this repo's models/nori_description/ is the source
  • Isaac Sim — isaac-mcp (NVIDIA Omniverse) USD import of the same nori_a3_posed.glb/URDF (isaac-mcp/scenes/nori_a3_sim.usda — 46 links, 45 joints, full rest-pose FK, oriented STL proxy collision boxes, and PhysicsDriveAPI, plus isaac-mcp/pipeline/08_nori_a3_active_inference_arm.py 7-DoF Active Inference controller via run_active_inference(robot="nori_a3")) — physics twin for sim2real

The result: open Resonite/Overte/Unity/Godot, call the fleet spawn tool, and the actual A3 geometry appears in-world — a virtual twin you can pose, inspect, or stage a pick-and-place task against before ever touching the real hardware. This repo supplies the mesh; the fleet does the spawning.

Quick Install

git clone https://github.com/sandraschi/norirobotics-mcp
cd norirobotics-mcp
uv sync
uv run norirobotics-mcp

See INSTALL.md for the drag-and-drop .mcpb path and full options.

Example Prompts

  • "What are the Nori A3's specs, and what did Hacker News say about it at launch?"
  • "Open a Nori session and check the current telemetry."
  • "Start an episode recording for pouring water into a cup, then stop it after I confirm."
  • "Register a physical A3 profile for the office unit and switch to it."
  • "Save this spot as a waypoint called 'kitchen', then navigate back to it."

Documentation

Doc Contents
Installation All install methods, prerequisites
Onboarding Mock-vs-real session, Supabase credentials, pre-hardware posture
Wrapped SDK What Nori A3 / nori-sdk is, lineage, community, disambiguation, the founder's background, and the founding arXiv paper's actual version history
The Affordable Home Robot Paradigm Why wheels-not-legs, a telescoping column, tiered end effectors, and split onboard/server compute look like the near-term future of home robotics — Nori A3 as the case study
Field Review Pros/cons/verdict — extols the wheels+telescoping-lift+open-SDK bet, and names the concrete gaps (no charging station, light payload, RC-grade actuators, no torso pitch to reach the back of a cabinet)
Actuator Alternatives Feetech vs. CubeMars/MyActuator/Damiao/RobStride QDD actuators — real specs/prices, why QDD gives force feedback nearly for free, torque-vs-speed for a homebot vs. a running dogbot, China sourcing/customs, and why "use a Unitree motor" isn't quite the right framing
DIY A3 Equivalent Full from-scratch build matching or exceeding every A3 spec — real torque math, telescoping-lift alternatives, 2-finger vs. 3-finger grippers, using this repo's own official (but non-watertight) mesh package plus blender-mcp/qcad-mcp to reconstruct real printable parts, costed Match/Exceed tiers
DIY Base Station Closing the no-charging-station gap — Roomba's real IR-homing + wiping-contact mechanism vs. a pogo-pin design (real current/spring-force/cycle-life specs), mechanical alignment, and what a wheeled base like A3's would actually need to add
DIY Project Plan The actual build sheet — BOM with named EU/Shenzhen vendors, maker tools, time-per-phase estimates. Honest up front: won't save much money, but it's a genuinely great build
Architecture Session model, ports, data flow
Configuration Env vars, config options
Tool Reference All 9 MCP tools
Development Contributing, local setup
Troubleshooting Common issues

Requirements

Python 3.11+, Claude Desktop (or any MCP client). No physical robot required for development — nori_sdk's mock session covers the full API surface.

License

MIT

from github.com/sandraschi/norirobotics-mcp

Installing Norirobotics

This server has no published package — it is built from source. Open the repository and follow its README.

▸ github.com/sandraschi/norirobotics-mcp

FAQ

Is Norirobotics MCP free?

Yes, Norirobotics MCP is free — one-click install via Unyly at no cost.

Does Norirobotics need an API key?

No, Norirobotics runs without API keys or environment variables.

Is Norirobotics hosted or self-hosted?

Self-hosted: the server runs locally on your machine via the install command above.

How do I install Norirobotics in Claude Desktop, Claude Code or Cursor?

Open Norirobotics on unyly.org, pick your client tab (Claude Desktop, Claude Code, Cursor) and press Install — the config is generated automatically, no JSON editing.

Related MCPs

Compare Norirobotics with

Not sure what to pick?

Find your stack in 60 seconds

Author?

Embed badge for your README

Browse similar

All development MCPs