Observability & data export

Every hosted inference session Reflex runs for you is logged, recorded, and metered, and all of it is pullable with just your API key — no dashboard, no export job. Three layers:

• Sessions — what ran: which model, runtime, status, timing.
• Recordings — the data: the full trajectory (states, actions, camera video) in LeRobot format.
• Usage — the cost: GPU time per deployment.

Everything is org-scoped: your key only ever returns your own data.

Sessions — what ran

import reflex

c = reflex.Client(api_key="rfx_...")

for s in c.sessions.list()["sessions"]:
    print(s["_id"], s["baseModel"], s["status"], s.get("runtime"), s.get("closeReason"))

Each session always has _id, baseModel, status, and createdAt. Depending on the run it may also carry runtime, robotType, mode, endedAt, and closeReason — use s.get("...") for those, since they're absent until set. Filter by status — authorized, active, closing, closed, expired, or failed:

active = c.sessions.list(status="active")["sessions"]

Recordings — the trajectory data

Every session is auto-recorded in LeRobot format (states, actions, camera video) and stored org-scoped. List them, then pull one into your own storage:

recs = c.recordings.list(limit=20)["recordings"]
for r in recs:
    print(r["id"], r["sessionId"], r["formats"], r["sizeBytes"], r["status"])

# stream the dataset to disk
path = c.recordings.download(recs[0]["id"], "session.lerobot.tar")

Each recording has id, sessionId, formats (e.g. ["lerobot"]), sizeBytes, sha256, status, and createdAt. download(id, dest) mints a presigned URL and streams the dataset to dest; download_url(id) returns just the signed link if you'd rather stream it yourself.

Usage — the cost

for receipt in c.receipts.list()["receipts"]:
    print(receipt["deploymentId"], receipt["createdAt"], receipt["receiptJson"])

Pass c.receipts.list(deployment_id="...") to scope to one deployment. Each receipt carries deploymentId, createdAt, and a receiptJson payload with the metered detail.

Correlating them into your database

The three layers join on the identifiers they return: a recording's sessionId matches its session's _id, and receipts carry deploymentId. A typical export pipeline:

sessions = {s["_id"]: s for s in c.sessions.list()["sessions"]}

for rec in c.recordings.list(limit=100)["recordings"]:
    session = sessions.get(rec["sessionId"])          # the run this trajectory came from
    c.recordings.download(rec["id"], f"data/{rec['sessionId']}.tar")
    save_to_your_db(session=session, recording=rec)   # your storage

API reference

All five take only your API key and return only your org's data.