The ClawDOS chat screen scrolls on a real IBM PC 5160, showing AI‑generated code and output.
*An AI chat harness now lives on a 1981 IBM PC XT, compiling and executing code in real time. The experiment tests Claude’s language model against the hard limits of 4.77 MHz hardware, exposing both nostalgic charm and modern security questions.*
An AI chat agent named ClawDOS is now running on a genuine IBM PC XT, the 8088‑based machine that launched the PC era. The 5160, clocked at 4.77 MHz with 640 KB of RAM and a CGA card, hosts a 176 KB DOS program that talks to Claude’s API through a separate bridge PC handling TLS. The harness writes, edits, compiles and executes code directly on the vintage hardware, keeping a persistent conversation log in C:\MEMORY. Creator Ricapar built the system for personal entertainment, aiming to push an AI coding agent to build something within the narrow limits of the era and then let that creation build further. Every interaction occurs on the native OS; the AI issues commands, the harness parses streamed replies, runs tools, and returns results, all without modern OS services. The result is a functional development environment that lives entirely on the original machine.
ClawDOS is a 176 KB DOS program (HARNESS.EXE) that speaks Claude’s Messages API. It runs on an IBM PC 5160 equipped with an 8088 CPU at 4.77 MHz, 640 KB RAM and a CGA card. The harness writes API requests to a bridge PC that handles TLS, then parses streamed replies back on the XT. It can read, edit, and write files, invoke compilers, run programs, and capture screenshots. All conversation state is stored in C:\MEMORY, allowing the AI to resume work where it left off. The system operates entirely on the native OS and hardware, with no modern OS services beyond the bridge for encryption.
In a demo request to "print the primes below 100," ClawDOS produced output 64 seconds after the user pressed Enter. Of that, 26 seconds were spent compiling, linking and running the program with Turbo C. The compiler runs with roughly 450 KB of free memory on the 8088, leaving enough headroom for the harness to resume the conversation. The timing shows that even with a 4.77 MHz processor, an LLM can drive a full build cycle without stalling the machine, albeit at human‑scale latency.
ClawDOS recreated a tribute to the 1983 film WarGames entirely on the XT. The project consists of five Turbo C source files totaling 15 KB, compiled into a WG.EXE under 10 KB. The demo draws a vector world map, animates missile arcs, and runs a tic‑tac‑toe endgame, outputting sound through the PC speaker or an AdLib card. A timer‑interrupt hook captures the CGA video memory 25 seconds into the run, encodes a 320 × 200, four‑colour PNG of 16 KB, and sends it back to Claude for analysis. The AI then edits the source to add an automatic exit, demonstrating a full edit‑compile‑run loop on vintage hardware.
The harness offers eighteen built‑in tools. File utilities include sysinfo, dir, read, search, write, edit, delete, mkdir and rename. Execution tools cover run (up to twelve commands per call with optional screenshot), peek, poke and port for memory and I/O access. Creative tools provide play and textart for PC‑speaker music and CP437 graphics. Memory tools include memory_list, memory_read and memory_write. Development tools comprise two assemblers (TINYASM and WASM), Bruce’s C, GW‑BASIC, the lDebug debugger and Turbo C 2.0, all installed from floppy images in an external directory. The AI can invoke any of these in a single tool call, allowing it to build, test and iterate code without leaving the XT.
ClawDOS offloads TLS to a separate bridge PC; the 8088 performs all other tasks while the bridge manages encrypted communication.
ClawDOS can invoke Turbo C 2.0, two assemblers (TINYASM and WASM), Bruce’s C, GW‑BASIC, the lDebug debugger, plus file and memory utilities.
In the prime‑printing demo, ClawDOS returned output 64 seconds after Enter, with 26 seconds spent compiling, linking and running.
The project blurs the line between retro computing hobby and cutting‑edge AI research. By forcing Claude to operate within the constraints of a 1980s PC, ClawDOS reveals both the robustness of modern language models and the lingering relevance of low‑level hardware knowledge. As AI continues to infiltrate every layer of software development, experiments like this warn that legacy systems may soon become unexpected front‑lines for security scrutiny and creative hacking.
Sources: Hacker News article 'ClawDOS – An AI coding agent for DOS' (http://ricapar.net/ClawDOS/)