reflex-sdk 0.7.4

Python SDK for Reflex hosted robot inference and training

Reflex Python SDK

Public Python package for hosted Reflex training, datasets, managed instances, and action inference.

python -m pip install reflex-sdk

The package installs the reflex Python module and a reflex CLI command in the active Python environment. If your Python installer puts console scripts in a directory that is not on PATH, use the module entry point instead:

python3 -m reflex --help

import reflex

This package is intentionally separate from inference/, which owns model serving and Prime worker runtime code.

Running a robot session (reflex connect)

reflex connect runs a closed-loop robot session (observe → infer → apply), fully described by a YAML config:

reflex login                       # or: export REFLEX_API_KEY=rfx_...
reflex connect --config robot.yaml

The config's target: block selects where inference runs (kind:):

kind	Where inference runs	Notes
webrtc	Hosted Reflex cloud (default)	Convex authorizes the session and routes it to a worker. A MolmoAct2 model (molmoact2-bimanualyam) is already hosted and online, so you can connect to cloud inference without deploying anything — just log in.
edge	Your own HTTP server on the LAN	Point url at a self-hosted inference server (e.g. a Jetson). No cloud/auth.

target:
  kind: webrtc          # hosted cloud inference (molmoact2-bimanualyam, already online)
  timeout_s: 60
  state_field: state
  instruction_field: prompt
  img_size: 256
hardware:
  kind: yam_bimanual
  config: { instruction: "fold the towel" }

Choosing the cloud provider — RunPod serverless

Hosted sessions (kind: webrtc) default to Modal. To request routing to a RunPod serverless node instead, add a runtime hint:

target:
  kind: webrtc
  runtime: runpod-serverless   # route this hosted session to RunPod (QUIC)

The RunPod account key stays server-side (same trust model as Modal) — you only need your Reflex API key. Convex routes to RunPod when a RunPod node is registered and available, otherwise the session uses the default backend. You can also set it via the environment: REFLEX_RUNTIME=runpod-serverless.

See examples/ for full configs (yam_bimanual_molmoact2_MODAL.yaml, etc.). To exercise reflex connect end to end without a robot (fake hardware + synthetic cameras, against the offline/Modal/RunPod backends), see sim/.