What is WLKATA MT4 control stack?
This project, from the GitHub user enghoff, replaces the software of the WLKATA MT4, a four-axis desktop robot arm, from end to end. It flashes new firmware onto the arm's ATmega2560 controller with Cartesian motion and queued paths, adds a fixed USB camera that maps printed ArUco markers to the arm's coordinates, and exposes 17 MCP tools so a client can home the arm, look at the desk, and pick up or place objects named by id, such as cube_2 or marker_1.
It is written for Python 3.10+ and runs on Windows. Optional GPU services add Grounding DINO for finding objects by name, SAM 2.1 for outlines and Qwen3-VL-4B for typed instructions. The repository has 1 star, commits from July to 15 August 2026, and no licence file. The arm itself costs $1,380 as an education kit at RobotShop.
How to install WLKATA MT4 control stack
pip install -r requirements.txt && python flash_jog.py --port COM6 && python -m mt4_mcp --stdioIs WLKATA MT4 control stack worth it?
The MT4 stack is serious engineering for one specific arm. Everything on the desk is addressed by entity id, and when the arm cannot pick or place something the tool says why: out of reach, inside the keep-out zone around the base, too little room for the fingers, off the desk. There is deliberately no nearest-match fallback, because nothing in the system could tell that the wrong object had been picked. That is the right design for an agent in charge of a physical machine.
This server moves a motorised arm with a gripper. The firmware enforces a 140 mm keep-out cylinder around the base, soft joint limits and a desk floor, but the README says joint-space moves and homing bypass those checks. Calibration matters as much as code: five scripts map the camera to the desk, and if a fresh base calibration is not followed by the rest, cube picks silently regain 15 to 30 mm of error and the desk-edge check stops refusing places past the edge. The J4 wrist zero is lost on every power cycle.
The catch: the network and the licence. By default the MCP server listens on 127.0.0.1:8787 or runs over stdio. A public mode for ChatGPT and other remote clients turns off host-header protection and pairs with an ngrok tunnel and Google sign-in; in our reading of auth.py there is no setting that limits which Google accounts may sign in, so a public tunnel should be treated as handing the arm to anyone who completes that sign-in. And with no licence file the code is all rights reserved by default: you may read it, but copying, modifying or redistributing it is not granted.
Our take: as a reference for giving an agent safe, explainable control of a robot arm, this is one of the more instructive repositories we have read, and its calibration and accuracy notes are unusually frank. As something to install, it suits a single MT4 owner on Windows who asks the author for permission first.
Where it earns its place
A robotics lecturer with one MT4 in the teaching lab
Show an agent turning a plain request into a safe pick-and-place
After calibration, connect Claude Code over stdio and ask it to put the red cube on marker 3; mt4_scene shows each entity's pickable flag and reason. The kit is $1,380; ask the author before reusing the code in course material.
A researcher prototyping language-guided manipulation
Test how a vision-language model's choices hold up against real reach limits
Run ask_qwen.py with --dry-run so Qwen3-VL-4B decides without moving, then compare its targets with the arm's own pickable verdicts. Needs a CUDA GPU for the Qwen and Grounding DINO services.
A hobbyist with an MT4 and a webcam
Have the arm stack cubes on a marker unattended
Run stack_cubes.py --marker 4 --preview, which clears the site, picks with a re-grip and builds up to 9 levels by default. Re-run calibrate_j4.py after every power cycle first.
Before you commit
- Setup
- pip install -r requirements.txt, flash the firmware with flash_jog.py (stock images are backed up and restorable), print the ArUco marker sheet, run the five calibration scripts, then start python -m mt4_mcp or register it over stdio in Cursor or Claude Code. Needs Windows, Python 3.10+ and PlatformIO with avrdude.
- Integrations
- WLKATA MT4 over a CH340 USB serial link at 115200 baud, a USB scene camera, optional Grounding DINO and Qwen3-VL-4B HTTP services, SAM 2.1 in-process, ngrok and FastMCP's Google OAuth for remote clients, and an Xbox gamepad for jogging.
- Data and privacy
- Camera frames and scene data go to your model provider through the client. Calibration is stored in vision_calibration.json. HTTP defaults to 127.0.0.1:8787; MT4_MCP_PUBLIC=1 disables host-origin protection for tunnels, and OAuth needs a Google client ID and secret.
- Cost at scale
- The code is free to download but unlicensed. The WLKATA MT4 Education Kit is $1,380 at RobotShop; the optional vision models want a CUDA GPU, though they run slowly on CPU.
- Limits
- One process owns the serial port at a time. The J4 zero is lost on power cycle or reflash. Skipping later calibration steps adds 15 to 30 mm of pick error without any failure. Payload is 500 g according to RobotShop.
What’s good
- Refuses with a reason instead of picking the wrong object
- Firmware keep-out zone, soft limits and desk floor
- Detailed calibration and accuracy documentation
- Optional open-vocabulary grounding by object name
What’s not
- No licence file, so reuse is not granted
- Windows only; needs the WLKATA MT4 arm and a camera
- Public mode has no account allowlist that we found
- Joint-space moves and homing bypass the keep-out checks
Best for
An owner of a WLKATA MT4 on Windows who wants an agent to move objects on a calibrated desk and is comfortable flashing custom firmware.
WLKATA MT4 control stack: common questions
- Is WLKATA MT4 control stack worth it?
- Careful, well-documented robotics work that needs the exact arm, a Windows PC and an afternoon of calibration, and that you are not legally free to reuse because it carries no licence. We rate it 3.2 out of 5.
- How much does WLKATA MT4 control stack cost?
- Free to download, but no licence is published. The WLKATA MT4 Education Kit costs $1,380 at RobotShop.
- What are the downsides of WLKATA MT4 control stack?
- No licence file, so reuse is not granted; windows only; needs the WLKATA MT4 arm and a camera; public mode has no account allowlist that we found; joint-space moves and homing bypass the keep-out checks.
- Who is WLKATA MT4 control stack best for?
- An owner of a WLKATA MT4 on Windows who wants an agent to move objects on a calibrated desk and is comfortable flashing custom firmware.
- How much does WLKATA MT4 control stack cost for a team or in production?
- The code is free to download but unlicensed. The WLKATA MT4 Education Kit is $1,380 at RobotShop; the optional vision models want a CUDA GPU, though they run slowly on CPU.
- Is WLKATA MT4 control stack safe for company data?
- Camera frames and scene data go to your model provider through the client. Calibration is stored in vision_calibration.json. HTTP defaults to 127.0.0.1:8787; MT4_MCP_PUBLIC=1 disables host-origin protection for tunnels, and OAuth needs a Google client ID and secret.
Answers checked Nov 14, 2026.