Developer preview · Documentation

Build one auditable robot Cell.

CenterOS gives you one versioned object model across heterogeneous hardware, teleoperation, recording and evaluation. This page describes what the preview packages do today.

Not on PyPI yet · configuration-specific

Request access Open the platform

Overview

One control and evidence contract, not one fake robot abstraction. CenterOS unifies identity, sessions, clocks, recording and evaluation. Vendor-specific kinematics, force modes, tactile layouts and safety states stay inside each device's adapter.

Packages 3 packages

Package	Version	Provides
`centeros-cli`	0.2.0	`centeros` command and the `centeros` Python SDK
`centeros-mcp`	0.1.0	`centeros-mcp` MCP server
`centeros-robotics`	0.1.0	`centeros-robot` command and `BaseTeleopAgent`

All three need Python 3.10 or newer and are MIT-licensed.

Install

The packages are not on PyPI yet. Preview participants receive them directly and install from the package path.

terminal

$ pip install ./centeros-cli
$ pip install ./centeros-mcp          # optional: AI agents
$ pip install ./centeros-robotics     # optional: robot agents

Authentication

Sign in once with the CLI. The token is saved to ~/.centeros/config.json and the Python SDK reuses it. Give the MCP server its token through CENTEROS_TOKEN.

terminal

$ centeros login             # browser sign-in (Google, GitHub or email)
$ centeros login --device    # device code, for headless hosts
$ centeros whoami
$ centeros pat --help        # personal access tokens for CI

Environment variables MCP server

Variable	Purpose
`CENTEROS_TOKEN`	Token for the MCP server (a login token or a personal access token)
`CENTEROS_API_URL`	Override the platform API URL (rarely needed)

CLI

Run centeros --help for every command group. The common ones:

terminal

$ centeros hardware list          # hardware catalog
$ centeros hardware specs wuji-hand
$ centeros teleop sessions        # recent teleop sessions
$ centeros teleop monitor <id>    # live event stream
$ centeros datasets list
$ centeros eval run --help        # start an evaluation run
$ centeros safety active          # active e-stops
$ centeros safety estop --help

All command groups centeros 0.2.0

Identity: login, logout, whoami, register, pat
Hardware and robots: hardware, robot, fleet, safety, ota
Sessions: teleop, sim, slam, mission
Data: datasets, dataset, share, sharing, cloud-import
Models: eval, training, models, vla
Platform: health, stats, actions, skill

MCP server

Add centeros-mcp to any MCP client and pass your token.

mcp.json

{
  "mcpServers": {
    "centeros": {
      "command": "centeros-mcp",
      "env": { "CENTEROS_TOKEN": "<your token>" }
    }
  }
}

Tool groups selection

Hardware: list_hardware_catalog, get_hardware_entry
Robots and fleet: list_robots, get_robot_twin, fleet_overview, list_fleet_alerts
Teleop: list_teleop_sessions, get_teleop_session, send_robot_command
Safety: get_safety_profile, list_active_estops, emergency_stop
Simulation, SLAM, missions: start_rc_sim, list_slam_sessions, list_missions
Data and docs: browse_datasets, search_wiki
Platform actions: list_available_actions, centeros_action

Motion gating send_robot_command

Commands are checked against the safety profile for that robot kind, velocity is clamped to profile limits (the agent is told), and commands are refused while an e-stop is active.

Python SDK

The SDK ships inside centeros-cli. It wraps the platform API with one client object.

quickstart.py

from centeros import CenterOSClient

client = CenterOSClient()          # or CenterOSClient(token="...")

status   = client.health.get()
datasets = client.datasets.list(limit=20)
robots   = client.fleet.robots()
models   = client.models.list()

Resources on the client CenterOSClient

datasets, training, models, fleet, annotations, stats, health, actions, orgs, sharing, review, feedback and more. API errors raise centeros.APIError; sign-in problems raise centeros.AuthError.

Robot agents

Write an adapter for your own robot by subclassing BaseTeleopAgent. The bundled PiPER, OpenArm and ALOHA agents are interfaces that run with --mock; hardware drivers are not included in the package.

terminal

$ centeros-robot list-robots
$ centeros-robot run piper --mock
$ centeros-robot run openarm --mock --record

Cell Runtime

On the edge host, a Cell manifest pins every component, and a session gate decides what may move.

Object model 5 records

Record	What it is
Asset	One physical device: serial number, model, firmware, driver version
Capability manifest	The device's support level, action and observation contract, and safety features
Cell manifest	Versioned set of components plus environment, calibration and data-contract versions
Session	Mode, local operator, state and command lease for one run
Evaluation run	Acceptance criteria and a verdict: pass, fail or inconclusive

Session modes 5 modes

Mode	Purpose
`calibration`	Calibrate the Cell
`teleop_capture`	Human-operated demonstrations with recording
`shadow`	Policy outputs observed, not applied to actuators
`canary_autonomous`	Restricted physical validation after a release gate
`autonomous`	Defined in the runtime; not offered in the preview

Session states lifecycle

created → active → halted or estopped → closed. Commands carry a lease (250 ms by default) and are rejected after it expires.

Support levels

Every device in the runtime catalog carries exactly one level.

Level	Meaning
Catalog	Model is known. No CenterOS driver; no control claim.
Observe only	Telemetry may be read, never commanded.
Experimental control	An adapter exists behind the local session gate. Not production-validated.
Certified control	Defined in the runtime. No device holds it today.

See the hardware matrix

Safety boundary

Control access is ours. Functional-safety certification is not.

CenterOS owns	Your native safety layer owns
Session orchestration and authorization	Hardware emergency stop
Adapter contracts and command leases	Safety PLC and certified interlocks
Recording, provenance and evaluation	Real-time torque and motion protection
Software e-stop requests	Site-specific risk assessment