pycentauri MCP server
Allows Home Assistant to view the printer's webcam stream as a camera entity via the RTSP bridge.
Enables Jellyfin to stream the printer's webcam feed as a live source via the RTSP bridge.
Allows Synology Surveillance Station to record the printer's webcam feed via the RTSP bridge.
Click on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@pycentauri MCP servercheck the printer status"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
pycentauri
Local-network toolkit for Elegoo Centauri Carbon 3D printers — the original Centauri Carbon (CC1) and the Centauri Carbon 2 (CC2). One async client, six surfaces: Python library, CLI, MCP server for AI agents, REST/SSE HTTP server, built-in web dashboard, and an RTSP bridge for your NVR.
No cloud account, no Elegoo servers — everything talks directly to the
printer on your LAN. pycentauri auto-detects which model it's talking
to and speaks the right protocol:
CC1 | CC2 | |
Transport | SDCP v3 over WebSocket ( | JSON-RPC over MQTT ( |
Auth | none | access code (printer screen) |
Discovery | UDP broadcast | direct IP + HTTP bootstrap |
Webcam | MJPEG | MJPEG |
Live head speed | — | ✓ ( |
Fan channels | 3 | 5 |
Canvas multi-filament | — | ✓ (status + auto-refill) |
Filament-switch detection | — | ✓ (position-based) |
Status: alpha, but used daily against real printers. Protocols were reverse-engineered from Elegoo's official
elegoo-linkC++ SDK, theCentauriLinkproject, and live wire captures. CC1 tested on firmware V1.1.46 and OpenCentauri V0.3.0-o; CC2 tested on firmware 01.03.02.51. Full wire-protocol notes live indocs/PROTOCOL.md.
Install
pip install pycentauri # library + CLI
pip install "pycentauri[mcp]" # + MCP server
pip install "pycentauri[server]" # + HTTP REST/SSE server + web UI
pip install "pycentauri[mcp,server]" # all Python surfacesThe RTSP bridge additionally requires
MediaMTX and ffmpeg
on $PATH.
Python 3.10+. Core dependencies: websockets, paho-mqtt, httpx,
typer, pydantic.
Related MCP server: SpaceTraders MCP Server
Connecting to your printer
CC1 needs only its IP (or nothing at all — it answers UDP discovery).
CC2 needs its IP and its access code, found on the printer's
touchscreen under network/connectivity settings. Pass it as
--access-code / access_code= / PYCENTAURI_ACCESS_CODE. The examples
below use Ab3dEf as a stand-in — substitute your own.
Enable "LAN Only" mode on the CC2 (network settings on the touchscreen). The CC2 gates its local API behind it — with LAN Only off the printer works through Elegoo's cloud and leaves the local HTTP endpoint closed, so pycentauri can't reach it and you'll get a connection error. This is required on firmware 2.0 and recommended on all CC2 firmware.
Every CLI command accepts --host (env: PYCENTAURI_HOST). With no host
given, commands try UDP discovery, which only finds CC1s.
CLI
# Discovery (CC1 only — CC2 doesn't answer broadcasts)
centauri discover
# Status, attributes, live watch, snapshot
centauri status --host 192.168.1.209 # CC1
centauri status --host 192.168.1.189 --access-code Ab3dEf # CC2
centauri status --host 192.168.1.209 --json
centauri attributes --host 192.168.1.209
centauri watch --host 192.168.1.209
centauri snapshot --host 192.168.1.209 shot.jpg
# Print control — all writes require --enable-control
centauri print start cube.gcode --host 192.168.1.209 --enable-control
centauri print pause --host 192.168.1.209 --enable-control
centauri print resume --host 192.168.1.209 --enable-control
centauri print stop --host 192.168.1.209 --enable-control
# Live adjust while printing
centauri speed sport --host 192.168.1.209 --enable-control
centauri fan --model 100 --aux 60 --chamber 30 --host 192.168.1.209 --enable-control
centauri temp --nozzle 215 --bed 60 --host 192.168.1.209 --enable-control
# Canvas multi-filament (CC2 only)
centauri canvas --host 192.168.1.189 --access-code Ab3dEf
centauri refill --on --host 192.168.1.189 --access-code Ab3dEf --enable-controlcentauri canvas prints each tray's filament, color, temperature range,
and loaded state:
auto_refill : OFF
active_tray : none
connected : yes
canvas #0:
● tray 0: PLA Wood (PLA) #F72221 [190-230°C]
● tray 1: PLA Wood (PLA) #AF7832 [190-230°C]
● tray 2: PETG (PETG) #A03BF7 [230-260°C]
● tray 3: PLA Wood (PLA) #D2C5A3 [190-230°C]Speed modes
Both printers accept exactly four speed settings — arbitrary percentages are silently ignored by the firmware:
Mode | CC1 wire value ( | CC2 wire value ( |
| 50 | 0 |
| 100 | 1 |
| 130 | 2 |
| 160 | 3 |
Speed changes only take effect while a print is actively running.
CC2 speed pinning (the firmware fights you, so pycentauri fights back)
The CC2 firmware resets the speed mode back to balanced on every Canvas filament switch. On a multi-color print that switch happens every couple of minutes, so set sport once and the printer quietly drops you to balanced at the next color change. This is firmware behavior; it happens whether you set the speed from pycentauri or from the printer's own touchscreen.
pycentauri works around it with pin-and-enforce (CC2 only, requires
--enable-control):
Whatever mode you select gets pinned. Set it via pycentauri (pins immediately) or on the touchscreen (pinned after it holds ~8 seconds — the firmware never resets to a non-balanced mode, so a sustained sport/ludicrous/silent must be a human).
A firmware reset is re-applied. The reset fires a few seconds before the head parks at the chute, so when the speed drops to balanced pycentauri waits to see what follows: a filament switch (the head parking) within ~12 seconds means it was the firmware, and the pinned mode is re-applied as soon as the switch completes.
A human's balanced is honored. If the speed drops to balanced and no switch follows within ~12 seconds, that was you tapping balanced on the touchscreen — the pin is released and balanced stays. (On the wire a human tap is byte-identical to a firmware reset; the presence or absence of the following switch is what tells them apart. Measured reset→park lead times cluster at 6–8 s, well inside the window.)
The pin clears when the print ends.
So both paths just work: set any speed from either the app or the printer and it sticks across filament switches, and you can always drop back to balanced from the touchscreen whenever you want.
The CC1 has none of this — its speed mode stays where you put it, so
set_print_speed is a plain one-shot there.
Python library
import asyncio
from pycentauri import Printer, CC2Printer, connect_auto
async def main():
# Explicit CC1
async with await Printer.connect("192.168.1.209") as printer:
st = await printer.status()
print(st.print_status, st.progress, st.temp_nozzle)
# Explicit CC2
async with await CC2Printer.connect("192.168.1.189", access_code="Ab3dEf") as printer:
st = await printer.status()
print(st.temp_nozzle, st.raw["_cc2"]["gcode_move_speed"]) # mm/min; ÷60 = screen's mm/s
canvas = await printer.canvas_status()
for unit in canvas.canvas_list:
for tray in unit.tray_list:
print(tray.tray_id, tray.filament_name, tray.filament_color)
# Auto-detect — port-probes :3030 vs :1883 and returns the right class
async with await connect_auto("192.168.1.189", access_code="Ab3dEf") as printer:
attrs = await printer.attributes()
print(attrs.machine_name, attrs.firmware_version)
asyncio.run(main())Both classes expose the same API: status(), attributes(), watch()
(async iterator of live status), snapshot(), start_print(), pause(),
resume(), stop(), set_print_speed(), set_fan_speed(),
set_temperatures(), canvas_status(), set_auto_refill(). Write
methods require enable_control=True at connect time and raise
ControlDisabledError otherwise. Canvas methods raise PrinterError on
CC1 (no Canvas support over SDCP).
CC2-only telemetry rides along in Status.raw["_cc2"]: the live head
speed (gcode_move_speed — the commanded speed of the current move in
mm/min; divide by 60 for the mm/s figure the printer's screen
shows), speed_mode, filament runout sensor state,
firmware-computed remaining_time_sec, machine_status/sub_status
raw codes, and external_device (camera / U-disk presence).
HTTP server + web UI
# Read-only, loopback only
centauri server --host 192.168.1.209
# Read + write + RTSP, on the LAN (put an authenticating proxy in front)
centauri server --host 192.168.1.209 --bind 0.0.0.0 --port 8787 \
--enable-control --rtsp
# CC2
centauri server --host 192.168.1.189 --access-code Ab3dEf \
--bind 0.0.0.0 --port 8787 --enable-controlThe server holds a single long-lived connection to the printer (WebSocket for CC1, MQTT for CC2) with automatic reconnect and exponential backoff — it will never exhaust CC1's 5-connection limit.
The web UI at /ui/ is a clean dark dashboard, mobile-friendly and
dependency-free (no CDN assets — works on an air-gapped LAN): live
webcam, job progress with layer/ETA, printer state, thermals, kinematics
(with live head speed on CC2), pause/resume/stop, speed-mode selector,
fan and heater sliders that hydrate from live values, a Canvas panel
with per-tray color swatches and an auto-refill toggle (CC2), and RTSP
bridge controls. Control panels only render when the server was started
with --enable-control.
Endpoints
Method | Path | Notes |
|
| Redirects to |
|
| Web dashboard |
|
| Health + version + connection state |
|
| Latest status (typed summary + full |
|
| Model, firmware, mainboard ID |
|
| Single JPEG frame |
|
| MJPEG proxy (drop into an |
|
| Server-Sent Events stream of status pushes |
|
| UDP LAN scan (finds CC1s) |
|
| Canvas state (CC2; |
|
| OpenAPI documentation |
|
|
|
|
| † |
|
|
|
|
|
|
|
|
|
|
|
|
|
| RTSP bridge state (when |
|
| Toggle the bridge (when |
† requires the server to be launched with --enable-control; otherwise
the route isn't registered at all.
Temperature writes are bounds-checked server-side: nozzle 0–300 °C, bed 0–110 °C, chamber 0–60 °C.
MCP server (AI agents)
Give Claude Code, Claude Desktop, Cursor, or any MCP client eyes and hands on your printer:
# Read-only (status, snapshot, attributes, discovery, canvas)
claude mcp add pycentauri --env PYCENTAURI_HOST=192.168.1.209 \
-- python -m pycentauri.mcp
# With control tools
claude mcp add pycentauri-cc2 \
--env PYCENTAURI_HOST=192.168.1.189 \
--env PYCENTAURI_ACCESS_CODE=Ab3dEf \
-- python -m pycentauri.mcp --enable-controlThe target host is pinned in the server's environment at spawn time — a prompt-injected agent cannot redirect commands to an arbitrary IP, because no tool takes a host parameter.
Tool | Availability | Description |
| always | State, temps, progress, layer, position, fans |
| always | Model, firmware, mainboard ID |
| always | Webcam frame as MCP image — the model sees the print |
| always | UDP LAN scan |
| always | Canvas trays, colors, auto-refill (CC2) |
|
| Start a file already on the printer |
|
| Job control |
|
|
|
|
| Any subset of model/aux/chamber, 0–100% |
|
| Any subset of nozzle/bed/chamber, °C |
|
| Canvas auto-refill toggle (CC2) |
Control tools aren't merely gated — without the flag they are never registered, so they don't appear in the model's tool list at all.
RTSP bridge
Re-streams the printer's MJPEG webcam as H.264/RTSP for clients that don't speak MJPEG — Home Assistant, Frigate, Jellyfin, Synology Surveillance, VLC:
# Standalone (foreground, Ctrl-C to stop)
centauri rtsp --host 192.168.1.209
# → rtsp://<this-host>:8554/printer
# Integrated with the HTTP server — adds a STREAM panel to the web UI
centauri server --host 192.168.1.209 --rtsp --bind 0.0.0.0MediaMTX only runs the ffmpeg transcode while a client is connected, so
idle cost is zero. Tunables: --fps, --bitrate, --preset, --path,
--port (standalone) or the --rtsp-* variants on centauri server.
The bridge picks the correct camera port for CC1 vs CC2 automatically.
Print status codes
print_status in the API and library uses the CC1 firmware's code
space, extended with three codes for CC2 Canvas operations:
Code | Meaning | Code | Meaning | |
0 | Idle | 13 | Printing | |
1 | Homing | 14 | Error | |
5 | Pausing | 15 | Leveling | |
6 | Paused | 16 | Preheating | |
7 | Stopping | 27 | Switching filament (CC2) | |
8 | Stopped | 28 | Filament load complete (CC2) | |
9 | Completed | 29 | Unloading filament (CC2) |
The full table (including CC1's resin-inherited codes) is in
docs/PROTOCOL.md.
On the CC2, mid-print Canvas filament switches are detected by head position: the firmware never fully leaves its "printing" state during a switch, but the head parks at the purge chute behind the bed (y ≥ 258 mm, physically outside the printable area) for the duration. pycentauri reports code 27 the entire time the head is parked there mid-print.
Safety model
Explicit opt-in for writes. Every surface requires
enable_control=True/--enable-controlbefore any state-changing command is possible. Read-only is the default everywhere.Bounds-checked heaters. The library refuses temperature targets outside nozzle 0–300 °C / bed 0–110 °C / chamber 0–60 °C even though the firmware might accept them.
Unauthenticated HTTP surface. The REST server has no auth of its own. Bind it to loopback (the default) or put an authenticating reverse proxy in front before exposing it beyond localhost.
You own unattended printing. An agent with control tools can pause, stop, or heat your printer. Leaving one unattended is your call.
Known firmware quirks
CC1 (original Centauri Carbon)
5 concurrent WebSocket slots, hard. The 6th connection gets HTTP 500
"too many client". Slots free on close. The CLI opens one per invocation; the HTTP/MCP servers hold exactly one long-lived slot.Paused/errored states don't push Attributes. Every SDCP command needs the printer's
MainboardID, which normally arrives in an Attributes push — but not while paused or errored. pycentauri pre-seeds it from UDP discovery on every connect. If you callPrinter.connect()on a paused printer without discovery, passmainboard_id=yourself.Unknown commands crash the firmware. A few unrecognised SDCP commands in quick succession kill the printer's
appdaemon — and any active print with it. Don't probe undocumented command codes against a printer that's doing something you care about.The push scheduler goes dormant at idle (and a reboot doesn't wake it). The firmware can enter a state where
Cmd 512subscribes are acknowledged but no status frame is ever pushed while the printer sits idle — persisting across reboots (verified 2026-07-05 on V0.3.0-o). Starting a print revives pushes at full rate. One-shotCmd 0requests always work, so pycentauri automatically falls back to polling when a subscribe goes quiet (~7 s updates at idle, full-rate pushes while printing). Clients that rely purely on subscribe pushes will hang forever on an idle printer in this state.
CC2 (Centauri Carbon 2)
No UDP discovery. The CC2 ignores broadcast probes; specify its IP explicitly. Give it a static DHCP lease — it doesn't register a hostname with most routers, so its address drifts otherwise.
Access code required for MQTT, passed as the password with username
elegoo. The HTTP bootstrap (/system/info) wants the same code as anX-Tokenquery parameter — it ignores the header form.The webcam is unauthenticated. MJPEG on
:8080(any path) is open to anyone on your LAN, access code or not. That's the firmware's choice, not ours.Rate limiting. Rapid-fire MQTT requests (3+ back-to-back) trip a cooldown of a few seconds during which the broker silently drops responses. pycentauri's polling cadence stays under it; your scripts should too.
The firmware resets the speed mode to balanced on every Canvas filament switch. pycentauri pins your chosen mode and re-applies it automatically, while still honoring a deliberate balanced from the touchscreen — see CC2 speed pinning above for how it tells the two apart.
Registrations expire without an app-level PING. The printer forgets a registered client after several quiet minutes and silently stops answering that session's requests — the MQTT connection itself stays up, so there's no error to catch. pycentauri sends the SDK's
{"type": "PING"}keepalive every 30 s to hold the registration; if you write your own client, you must too.File list (method 1044) and video stream (1042) don't respond on firmware 01.03.02.51, so remote print-start on CC2 requires knowing the filename in advance.
Project layout & docs
src/pycentauri/
├── client.py # CC1: async SDCP-over-WebSocket client
├── cc2.py # CC2: async JSON-RPC-over-MQTT client (same API)
├── connect.py # connect_auto() — port-probe model detection
├── sdcp.py # SDCP v3 envelope build/parse
├── discovery.py # UDP broadcast discovery
├── camera.py # MJPEG frame grabber
├── models.py # Status / Attributes / CanvasStatus / PrintInfo
├── cli.py # Typer CLI
├── server.py # FastAPI app + connection supervisor
├── rtsp.py # MediaMTX/ffmpeg bridge
├── mcp/ # FastMCP stdio server
└── web/ # Static dashboard (no build step, no CDN)docs/PROTOCOL.md— both wire protocols in detail: envelopes, command/method tables with tested-on dates, status payloads, error codes, failure modes, and a CC1-vs-CC2 comparison.docs/ARCHITECTURE.md— module map and request flow.
Development
git clone https://github.com/bjan/pycentauri && cd pycentauri
python -m venv .venv && .venv/bin/pip install -e ".[mcp,server,dev]"
.venv/bin/ruff check . && .venv/bin/ruff format --check .
.venv/bin/mypy src # strict mode
.venv/bin/pytest -q # no printer required — tests use in-process fakesTests run against an in-process fake SDCP WebSocket server plus pure
translation-layer tests for CC2; nothing in CI touches real hardware.
Live verification against a physical printer is manual — centauri status, centauri canvas, and a fan write are the standard smoke
test after protocol-layer changes.
Credits & license
Protocol references: Elegoo's
elegoo-linkSDK (Apache-2.0) andCentauriLink.Licensed under Apache-2.0 — see LICENSE.
Not affiliated with or endorsed by Elegoo. Reverse-engineered protocols can break with any firmware update; nothing here is warranted to keep your prints alive.
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Latest Blog Posts
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/bjan/pycentauri'
If you have feedback or need assistance with the MCP directory API, please join our Discord server