This chapter is for developers who connect field devices (edge PCs) to MLOps Central Platform. You use the public PyPI package geo-mlops-sdk (import name geo_mlops_sdk, currently 0.3.0, Python 3.10 or later).

Two things in the package

WhatimportWhen
Agent (geo-mlops-edge)Not used — run it from a config fileRuns as an always-on service on the device. Has everything: local queue, offline buffering, uploads, collectors, model cache, local API
Client CentralClientgeo_mlops_sdk.clientWhen your existing application calls the central server API directly

In most cases the agent alone is enough. You configure it through edge.yaml and environment variables, with no code.

Flow between the device and Central

The link between the device and Central: the Edge SDK agent collects data from the robot arm, PLC and camera in the factory and keeps it in a local queue. The agent connects to Central first to register, send heartbeats, check commands and sync, and Central answers with commands and model files on that connection. A connection from Central into the factory stops at the firewall

What the edge device and the central server exchange. Every request starts from the device.

Calls go in one direction only: from the device to the server. The server cannot open a connection to a device behind a factory NAT, so the device also fetches commands by long polling (keeping a request open until a response arrives). Every call except the health check (GET /api/v1/health) is authenticated with a device token issued by Central (the X-Edge-Token header).

Pages in this chapter

PageContents
Installpip install and extras
5-minute startRegister a device → edge.yaml → run the agent → check it on the fleet screen
Configuration referenceEvery EdgeSettings field, config file lookup order, environment variables
Data collectorshttp · push · modbus · watchdir · frames, custom collectors
Models and inferenceModel download, runners, the local inference API
Local APIThe HTTP API on the device (port 8600)
Commands, policy and retentionServer commands, the Policy tab, offline buffering and sync
Connecting directly with CentralClientAsync client, error hierarchy, retries
Deploysystemd, Docker, docker compose, token rotation
CLIgeo-mlops-edge subcommands
Example: YOLO video demoAn app that runs YOLO on a video with the SDK runtime embedded: inference results, frame collection, model swaps
TroubleshootingOffline, 401, 403, 413, queue backlog

The full function signatures are in the Edge SDK API reference.

Written for the platform as of 2026-10-02.

© MLOps Central Platform