An automobile enthusiast achieved a breakthrough by utilizing an AI coding assistant to interface with the engine computer of a 1994 Jeep Cherokee. This project, which spanned approximately a week, was accomplished using a cost-effective $3 Arduino Nano. Remarkably, the vehicle successfully passed California’s stringent emissions test, as detailed by Muhammad Zulali in a YouTube video uploaded on November 10, 2026.
Challenges with Vintage Vehicle Diagnostics
The absence of a modern OBD2 port in the Jeep posed significant hurdles for conventional diagnostic tools. The vehicle initially failed the emissions test at 25 mph due to elevated nitrogen oxides (NOx) levels, a common issue when engines operate with a lean fuel mixture. Zulali’s initial attempt to rectify the issue by replacing the oxygen sensor resulted in even poorer performance at 15 mph, prompting him to seek a more innovative solution through Claude Code.
Claude Code identified that the Jeep’s engine computer used an outdated Chrysler communication protocol known as SCI, which required signal inversion compared to standard serial connections. A simple $3 Arduino Nano, configured on a breadboard, served as a translator, with the total component cost amounting to approximately $15.
Decoding the Jeep’s Engine Computer
According to reports from Tom’s Hardware, the diagnostic connection operated at a baud rate of approximately 7,800, utilizing only three wires from the vehicle’s diagnostic plug. Initial data readings at 1:30 a.m. were limited to coolant temperature and battery voltage. Subsequent data revealed peculiar fuel adjustments, showing a 22 percent increase in fuel trim, which led Zulali to investigate fuel pressure, discovering it was alarmingly low at 10 PSI.
Claude Code then demonstrated its capability to control vehicle functions by independently activating the fuel pump for brief intervals. However, a mistake was identified in the ChryslerScanner parameter table, which incorrectly indicated fuel addition rather than removal. This error was attributed to the scanner’s author, László Dániel, who was acknowledged in the project for his parameter contributions.
Innovative Solutions and Open Source Development
Claude Code proceeded to duplicate the engine computer’s program onto a laptop, revealing a California-specific fuel table that reduced fuel at the load used in emissions testing, explaining the high NOx levels. Although reflashing the engine control unit (ECU) seemed a logical fix, the ECU’s age precluded this option. OpenAI’s Codex provided a workaround by transmitting a compact program over the same wires, which was initially erased but later successfully stored in EEPROM, a type of memory that retains information even without power.
The entire project is available as an open-source initiative under the GPLv3 license, hosted on GitHub as the xj-scanner repository, which had garnered 252 stars and 30 forks at the time of documentation. The project was tested solely on a single vehicle model, the 1994 Cherokee XJ with a 4.0L engine, and users are cautioned about potential risks.
Implications and Security Considerations
This endeavor underscores the rapid adaptability of AI in understanding obsolete systems, a concern for cybersecurity experts who have previously demonstrated the vulnerabilities of vehicle computers to physical attacks, such as disabling brakes. AI agents now significantly reduce the expertise and time required for such interventions, presenting both opportunities and risks.
As highlighted by previous incidents where AI agents were manipulated, it is crucial for users to adhere to security protocols. Anthropic’s Claude Code security guide advises careful review of commands prior to execution. Zulali concedes that he lacks formal mechanical training, and some of his findings may require independent verification.
