plexus-python

Plexus is storage and dashboards for hardware teams: stream data from drones, robots and IoT devices into Plexus Time Series, or connect the database you already run, and get live dashboards and alerts. Website: plexus.company. Docs: docs.plexus.company.

This is the thin Python SDK for Plexus. Send telemetry to the Plexus gateway in one line. Storage, dashboards, alerts, and fleet management live in Plexus — this package just ships your data.

Get started

There is one way to set up Plexus: the Plexus MCP server, used from your coding agent.

With a coding agent (Claude Code, Cursor, Codex)

Add the Plexus MCP server. In Claude Code:

claude mcp add --transport http plexus https://app.plexus.company/mcp

Then tell your agent:

set up Plexus

Approve the browser sign-in once. The agent creates a send-only API key, sends one reading, and gives you the link to your dashboard. After that, ask it to change dashboards the same way.

On a device with no agent

pip install plexus-python
export PLEXUS_API_KEY=plx_xxx

Create the key on the API Keys page: app.plexus.company/api.

from plexus import Plexus

px = Plexus(source_id="device-001")   # reads PLEXUS_API_KEY
px.send("temperature", 72.5)

Device identity

Every device needs a unique source_id. Pass it to Plexus(source_id=...). It must match ^[a-z0-9][a-z0-9._-]*$ (max 256 chars).

Without source_id=... in code, the SDK makes up a random id like source-1a2b3c4d on first run and saves it to ~/.plexus/config.json. Don't use the hostname: cloned SD-card images all boot as raspberrypi, and their telemetry merges into one source.

Names are not auto-deduplicated. The gateway echoes back whatever source_id you declare, unchanged. Two devices that declare the same name write into the same source, so pick a unique one per device.

Core methods

send(metric, value) — stream a reading

The main method. Call it every time you have a new sensor reading.

px = Plexus(source_id="rig-01")   # reads PLEXUS_API_KEY from env

px.send("engine.rpm", 3450)
px.send("coolant.temp", 82.3)

metric is a dot-namespaced string ("motor.rpm", "gps.fix_quality"). value accepts any JSON-serializable type:

Optional arguments:

• tags={"motor_id": "A1"} — key-value labels for filtering in the dashboard
• timestamp=t — explicit Unix timestamp in seconds; omit to let the SDK pick (see Timestamps)

send_batch(points) — send multiple readings at once

Use this when you sample several sensors together and want them to share a timestamp and land in one network call.

px.send_batch([
    ("temperature", 22.4),
    ("humidity",    58.1),
    ("pressure",    1013.2),
])

points is a list of (metric, value) tuples, or (metric, value, timestamp) 3-tuples when you need a per-point timestamp. Points without their own timestamp share the batch timestamp (now, unless you pass timestamp=t).

batch() — coalesce a fast stream of readings

Use this above a few readings per second. Every send() is one WebSocket message, and the gateway limits messages, not points — 2,000/s on a connection. 25 channels at 100 Hz sent one at a time is 2,500 messages/s, and the overflow is discarded before it is stored.

with px.batch(interval_ms=50) as b:
    while running:
        b.send("att.pos_x", att.x)
        b.send("att.rate_x", gyro.x)
        b.send("frames.captured", grabber.count)

b.send() takes the same arguments as px.send(). A background thread flushes the queue every interval_ms, and leaving the block flushes what is left, so nothing is stranded. Readings keep the timestamp they were taken at, not the one they were flushed at.

What happens over the limit depends on which limit and which transport:

• WebSocket, over 2,000 messages/s on the connection: the gateway discards the message and sends a RATE_LIMITED error frame. The SDK counts those on px.rate_limited_frames and raises RateLimitedError on a later send. The discarded points are gone.
• WebSocket, over the per-source ceiling (2,000 messages/s, bursts up to 500): the gateway discards the message and sends nothing back. The SDK cannot see this loss.
• HTTP (the fallback, and the only path on the Free plan): the gateway answers 200 with a dropped count in the body. Since 0.14.0 the SDK counts those points (px.dropped_points) and the next send raises RateLimitedError, the same as on the WebSocket. Before 0.14.0 they were lost without a sign.

So RateLimitedError tells you about some drops, not all of them. Batch so you stay far under the ceiling.

run(name) — mark a test run

A run is a named window on a source. Runs are recalled on /runs, compared against each other aligned at T+0, and checked against their pass criteria when they close.

Runs need a paid plan. On the Free plan start_run() (and so px.run()) raises PlexusError with a 402.

with px.run("hotfire-03", pass_criteria=[
    {"metric": "motor.temp_c", "operator": "<", "value": 85},
]) as run:
    bench.execute()

Leaving the block closes the run as completed; an exception closes it as aborted and re-raises. Use px.start_run() / px.end_run() when the two halves happen in different places — end_run() returns the run with its verdict in test_result.

event(name, data) — record a discrete occurrence

Use event() for things that happen rather than things you measure continuously. Faults, state transitions, operator actions, log entries — anything you'd put on a timeline as a marker rather than plot as a graph.

px.event("fault",        "E-stop triggered")
px.event("state_change", {"from": "IDLE", "to": "RUNNING"})
px.event("sensor_error", {"sensor": "imu", "code": 42}, tags={"motor": "A"})

Limits per event: a string value up to 256 bytes, a dict or list value up to 4,096 bytes of JSON, and up to 16 tags. The gateway rejects anything larger.

Logs

There is no log-file upload and no logging.Handler in this package. To get important log lines into Plexus, send them as events:

px.event("log", {"level": "error", "msg": "IMU read timed out"})

Forward the lines you would want on the timeline next to your telemetry (errors, warnings, state changes), not every debug line. Each call is one message, and the gateway limits messages (see batch()).

Video streaming

Two methods depending on whether you control the capture loop or just have a URL.

Video needs a paid plan: frames go over the WebSocket, which the gateway refuses on the Free plan. Frames are relayed live to anyone watching. They are stored only when someone presses Record in the app, for up to 4 hours per recording.

send_video_frame(frame, camera_id) — send frames you capture yourself

Use this when your code owns the capture loop — a picamera2 callback, an OpenCV VideoCapture loop, or an FFmpeg pipe you manage. Pass each frame and the SDK ships it to Plexus over WebSocket.

import cv2

cap = cv2.VideoCapture(0)
while True:
    ok, frame = cap.read()
    if ok:
        px.send_video_frame(frame, camera_id="front")

Accepted frame types:

• numpy ndarray (H × W × C) — from OpenCV or picamera2; requires opencv-python
• JPEG bytes — passed through as-is, zero re-encode overhead
• Other image bytes (PNG, BMP, WebP) — decoded and re-encoded as JPEG via Pillow; requires pip install plexus-python[video]

camera_id identifies which camera the frame came from. Use distinct IDs when streaming from multiple cameras simultaneously ("front", "rear", "cam:0").

stream_camera(url, camera_id) — stream from an RTSP URL or file

Use this when you have an RTSP stream or video file and don't want to manage the capture loop yourself. The SDK runs FFmpeg internally and handles the rest. Requires FFmpeg on $PATH.

stop = px.stream_camera("rtsp://192.168.1.100/stream", camera_id="front")
# ... do other work ...
stop.set()  # stop streaming

Returns a threading.Event — call .set() to stop. Runs in a background thread so it doesn't block your main loop.

Which to use: if you're piping from rpicam-vid, picamera2, or your own capture process, use send_video_frame(). If you have an RTSP URL or file path, use stream_camera().

Bring Your Own Protocol

This package ships no adapters, auto-detection, or daemons — just the client. Use whatever library you'd use anyway and pipe values into px.send().

# MAVLink (pymavlink)
for msg in conn:
    if msg.get_type() == "ATTITUDE":
        px.send("attitude.roll", msg.roll)

# CAN (python-can)
for msg in bus:
    px.send(f"can.0x{msg.arbitration_id:x}", int.from_bytes(msg.data, "big"))

# MQTT (paho-mqtt)
def on_message(_c, _u, msg):
    px.send(msg.topic.replace("/", "."), float(msg.payload))

# I2C sensor (Adafruit CircuitPython)
px.send("temperature", bme.temperature)

See examples/ for runnable versions of each.

Reliability

A send goes straight to the gateway. If it fails, the points are saved to a local buffer, the error is raised, and the next send (or px.flush_buffer()) tries them again first. The HTTP path retries with exponential backoff (3 retries by default) before it gives up. The buffer is on disk (SQLite) by default, so it survives restarts and power loss. To keep it in memory only:

px = Plexus(persistent_buffer=False)

Point counts and flush:

px.buffer_size()
px.flush_buffer()

Timestamps and clock correction

By default — px.send("temp", 72.5) with no timestamp argument — the SDK picks the time itself. Over WebSocket, it synchronizes with the gateway clock on every connection, so data lands at the right place on the timeline even if the device's system clock is wrong (no NTP on first boot, stale RTC, fresh OS image).

px.send("temperature", 72.5)                # SDK picks time; gateway-synced over WS
px.send("temperature", 72.5, timestamp=t)   # your timestamp, used as-is, no correction

Pass an explicit timestamp when you have a reliable external time source (GPS, trusted RTC, host NTP) or are replaying historical data with known timestamps.

Omit timestamp when the device may have booted without NTP — which is the default on Raspberry Pi, Jetson, and most embedded Linux boards without a network connection at first boot.

Known limits:

• Clock sync refreshes on WebSocket (re)connect. A device with a drifting RTC that stays connected for many days accumulates uncorrected drift between reconnects.
• The HTTP fallback path (used when the WebSocket is unavailable) does not receive clock sync — timestamps default to the uncorrected device clock.
• send_batch() shares one timestamp across the batch by default; pass (metric, value, timestamp) 3-tuples for per-point timestamps.

Transport

By default the SDK connects over a WebSocket to /ws/device on the gateway — the gateway's device wire protocol. This gives you:

• lower-latency streaming of telemetry,
• the channel that carries commands from Plexus to your device.

If the socket is unavailable, sends transparently fall back to POST /ingest so no data is lost.

# ws with transparent http fallback — this is the only mode
px = Plexus()

There is no transport selector: the SDK always prefers the WebSocket and falls back to POST /ingest on its own when the socket is unavailable.

Either way, plain px.send() is one message per call; it does not batch. px.send_batch() sends one list as one message, and px.batch() groups a fast stream for you in the background.

On the Free plan the gateway refuses the device WebSocket (streaming_requires_plan). The SDK falls back to HTTP by itself, so send(), send_batch(), batch() and event() all still work. Live streaming, video, commands and runs need a paid plan. Free also caps you at 3 devices and 7 days of history.

Commands

Declare what your code can be asked to do. Declare before the first send(): the declaration travels in the auth frame.

from plexus import Plexus

px = Plexus(source_id="pod-07")

@px.command("power_off", title="Power off", danger="critical", idempotent=True,
            expires_in=30,
            params={"outlet": {"type": "integer", "minimum": 1, "maximum": 8}})
def power_off(run, outlet):
    pdu.outlet(outlet).off()                 # your code
    return {"outlet": outlet, "state": "off"}

px.serve()   # blocks until Ctrl+C / SIGTERM; or keep calling px.send(...)

Commands need a paid plan. They travel over the device WebSocket, which the gateway refuses on the Free plan, so on Free a handler is never called and px.serve() just waits.

Use an API key created with Receive commands on the Plexus API Keys page. Other keys keep sending telemetry, but their commands are ignored.

The handler is called as handler(run, **params), with every parameter already checked and coerced; a bad parameter is refused before your code runs. The return value becomes the run's result; an exception makes the run failed.

The SDK acknowledges each run, never runs the same run id twice, judges expiry on a monotonic clock, and replays unacknowledged statuses after a reconnect. The return value and any error text are sent to Plexus with the run's status. For an org set to keep command metadata only, Plexus discards them on arrival instead of storing them.

Run them from the Commands page or a dashboard panel in Plexus.

px.on_command(name, handler, ...) is deprecated. It still answers commands with its old handler(command_name, params_dict) signature, but is no longer advertised in the auth frame.

Environment Variables

The SDK also reads api_key from ~/.plexus/config.json if that file exists and PLEXUS_API_KEY is not set. Machines set up with an older version keep working.

Architecture

Your code ── px.send() ── WebSocket /ws/device (or HTTP POST /ingest) ──> plexus-gateway ──> ClickHouse + Dashboard

One thin path. No daemon, no adapters. Dashboards, alerts and fleet views are in the web app at app.plexus.company. This package gets your data there.

License

Apache 2.0