Keep the printer. Reuse the phone. Learn from real prints.
KlipperLearn combines a touch-friendly phone interface, local camera and sensor evidence, reproducible calibration charts, and bounded tuning proposals for compatible Klipper / Moonraker printers. Its purpose is to make existing machines more useful rather than require a new printer.
0.6.0 — file-first Printer Optimizer and optional MCP tools. This repository now contains the original Python application and its HTML/JavaScript phone companion, not only the earlier advisory demonstration. The original installation is not overwritten by this publication. Read the capability matrix before connecting hardware.
New preview: KlipperLearn Phone, a phone-only web app. Open https://agnuxo1.github.io/KlipperLearn/app/ in Chrome for Android, connect the printer's USB port with an OTG cable and tune Marlin-compatible printers without a Raspberry Pi or PC. It controls the printer over WebUSB, generates calibration prints, measures with the camera, microphone and accelerometer, and recommends one bounded change per trial. It has been tested only against a simulated printer so far. Printers flashed with Klipper firmware still need a Klipper host (the Python deployment below). No trained defect weights or universally validated automatic calibration profile are bundled.
Project website · Start here · Versioned downloads · Browser-only reference demo
The site and packages are free. The browser demo is the original 0.1.0 offline reference reviewer, not a hosted printer controller or the current three-mode Python optimizer. The ChatGPT directory listing and official upstream acceptance are still separate pending steps. Community status.
New in 0.6.0: a portable ChatGPT/Codex skill package
with a standard-library helper, optional MCP tools, original calibration coupon,
and an authenticated local workflow at /mobile/optimizer.html.
Review actual original photos and logs in your selected assistant. Propose one bounded experiment, repeat it, and export Quality / Standard / Speed as three matching process + filament pairs. The implemented native exporter targets OrcaSlicer; other slicers receive a reviewed settings report, not a fake native conversion. A phone, printer firmware change and cloud API call are not required for the file workflow.
Profiles are selected from whole comparable reviewed configurations, with a quality floor and at least two independent repetitions. Layer height/count, geometry, infill, supports, temperature and custom G-code are not changed. Missing evidence blocks output. Multiple modes can legitimately share the same settings. These are evidence-supported candidates, not universally optimal profiles.
Optional discovery identifies Moonraker/OctoPrint APIs on an explicitly approved private subnet, with strict scope and time limits. It does not scan Wi-Fi passwords, guess a printer model, start a print, or expose a private printer to the Internet. The package is not yet a registered/approved public ChatGPT plugin, and native slicer import and real physical results require separate acceptance tests.
Free and open source: no publisher subscription, external account or API key is required for the file-first plugin. ChatGPT eligibility and usage limits remain controlled by OpenAI. Public submission kit and current blocker.
Read the slicer workflow and plugin/MCP connection instructions.
| Mode | Available now | Deliberate boundary |
|---|---|---|
| Local learning | Local evidence storage, registered chart analysis, small ridge-model experiments, bounded parameter proposals, and optional MobileNet training/inference tools. | A fitted model is not proof of general print-quality accuracy. No pretrained defect model is bundled; low-quality or incomplete evidence must be rejected. |
| External advisor | An offline HTML reviewer exports a structured request and validates manually returned JSON from ChatGPT or another reviewer. | No cloud account, API key or automatic API call is built in. Returned text is not executable G-code and does not directly control a printer. |
Concept art: the diagrams describe the project direction, not a validated native-phone deployment.
Both modes use explicit evidence and bounded proposals. Hardware protection stays with Klipper and a qualified operator; neither mode may bypass thermal limits, trusted configuration or confirmation requirements.
- Touch console: printer status, temperatures, file selection, explicit manual controls, camera preview, history and human ratings.
- Phone evidence: motion, orientation, aggregated acoustic features, a durable upload queue, and controlled paired photographs when supported by the device. Raw microphone recordings are not saved by the sensor collector.
- Calibration workflow: chart generation and geometry checks, immutable file hashes, reviewed observations, one-variable proposals, and experimental six-zone screening. Specified limits are not a guarantee that a machine can safely reach them.
- Integrations: Moonraker APIs, camera registration/proxy interoperability with Mainsail, candidate Orca profile exports, and a source-preserving Orca slice manifest.
- Privacy: local storage by default, authenticated private routes, bounded request bodies, and no publication of the author's private keys, runtime data or old Git history.
Use Python 3.11 or newer on a compatible host. These steps install the application, not Klipper firmware, a native Android host, or printer-specific configuration.
git clone https://github.com/Agnuxo1/KlipperLearn.git
cd KlipperLearn
python -m venv .venvActivate the environment with .venv\Scripts\activate on Windows or
source .venv/bin/activate on Linux, then:
python -m pip install -e ".[learning]"
python -m klipperlearn serve --host 127.0.0.1 --port 8765Open http://127.0.0.1:8765/mobile/console.html. This default invocation does not enable printer control or connect to a printer. For an authenticated phone connection over HTTPS, follow the complete installation guide. Do not expose Moonraker or this application directly to the public Internet.
No desktop-PC dependency: install the Python backend and its private state on your Linux printer host, enable its system service, and open that host on the phone. Changing only the Moonraker address does not relocate a Windows backend. Follow the Linux deployment and migration guide, including HTTPS, startup, camera-proxy and instance-identity checks.
For the separate, dependency-free external-advisor reviewer:
python -m http.server 8080 --bind 127.0.0.1 --directory referenceOpen http://127.0.0.1:8080. Its clearly labelled synthetic example is for software demonstration and is not a printer profile. The reference page has no printer transport and its content security policy disables network API calls.
The intended loop is establish a baseline → collect synchronized evidence →
review → propose one bounded change → run an explicitly approved trial → compare
and repeat or stop. A failed trial remains useful negative evidence; a completed
Moonraker job does not prove dimensional accuracy or acceptable surface quality.
Automatic trial preparation must be explicitly enabled. Starting a new physical print still requires an explicit request and a clear bed; this release does not implement robotic bed clearing or a validated unattended safety supervisor. See the calibration protocol and local learning guide.
Attribution: these Benchy trials were printed on the Flashforge Creator Pro with operator adjustments advised by ChatGPT, not by KlipperLearn. The Anycubic 4Max Pro is the separate KlipperLearn + phone + webcam installation, using our charts and the Anycubic calibration cube. These are distinct machines, benchmarks and workflows; see benchmark provenance.
The artwork shows the author-reported tuning progression and is not an independent quality measurement. A subsequent review of the available printer history found a fastest identifiable completed, non-partial Benchy entry of 22 min 17 s print duration (25 min 35 s total duration), recorded with 0.4 mm layers and 121 layers. Other trials use different layer heights and counts. Printing a full-height model is not the same as preserving a fixed layer count or complying with a standardized benchmark protocol. The image is AI-retouched editorial artwork; its approximately 20-minute label must not be treated as the exact logged result. The original source-photo pixels and image provenance notes are available separately. Raw private histories are not published, and these times do not establish visual quality.
The current application separates the phone browser, Python companion, Moonraker, Klipper host, and printer MCU. Keep those roles distinct when reporting results. See architecture, Android host requirements, integration boundaries, model documentation, safety, and privacy.
These are external integrations, not changes accepted by Klipper, Moonraker, Mainsail, OrcaSlicer or Voron. Contributions should solve a concrete problem and follow the relevant project's process; see community collaboration.
python -m pip install -e ".[learning,dev]"
npm ci
python tools/run_offline_tests.py
node scripts/run-node-tests.cjs
node --test tests/core.test.js
python -m unittest discover -s tools/exporter/tests -v
python tools/check_repository.pyBrowser tests require Playwright Chromium: run npx playwright install chromium.
On Windows an already installed Edge can be selected with
KL_TEST_BROWSER_CHANNEL=msedge. All browser printer endpoints are simulated;
software tests never prove hardware safety. Details and measured results are in
validation and the 0.5.1 review.
The imported application is GPL-3.0-or-later; see LICENSE and COPYING. The separately published 0.1.0 reference reviewer and integration tools retain their MIT license in LICENSES/MIT-reference.txt. Third-party notices and the included Klipper logo are documented in THIRD_PARTY_NOTICES.md.
Created by Francisco Angulo de Lafuente and KlipperLearn contributors. Independent project: no upstream affiliation, endorsement or hardware certification is implied. Source-import provenance explains what was imported, revised, excluded and preserved privately.
The ten-project integration workstream tracks external adapters, contribution rules, tests and upstream acceptance separately. A PrusaSlicer post-processing setup reuses the existing read-only fingerprint helper. This is interoperability maintained here, not a claim that external projects have bundled or endorsed KlipperLearn.
KlipperLearn now provides project-specific external artifacts for Klipper, Moonraker, OrcaSlicer, Mainsail, Fluidd, OctoPrint, PrusaSlicer, Cura, KIAUH and Voron. This is interoperability work in our repository, not a claim that those ten projects accepted or included KlipperLearn.
Start with the downloadable packages or the project-by-project status and validation. The tools preserve printer/workflow provenance, avoid modifying G-code and make native-GUI, hardware and maintainer acceptance explicit rather than assumed.
The model-platform kit contains the original STL, an editable exact-output generator, a labelled geometry render, a reproducible validation report and platform-specific publication checklists. The complete application and versioned downloads remain available through the links above.
Current six-platform status distinguishes prepared materials, a saved Printables draft, actual public publication and acceptance. A draft or upload attempt is not public adoption. Physical validation of this exact coupon is not asserted.