Ken Barry · Cork · trading systems since 2012 · independent since April 2023 · this since August 2025
AuToken turns a sentence into Java, compiles it and runs it inside itself against sixty-three typed runners, and then a second model that never saw the request has to agree the result is right. Ask it for Hong Kong at night and it builds the city, the Peak behind it and the searchlights over it, in its own 3D world. It also talks, draws, writes music, and drives a phone, a printer and a rover.
3D worlds · built by a sentence
Scenes in AuToken's own OpenGL world. No model file, no imported scene, no texture loaded from disk: every object below was placed by a Java program a model wrote against the 3D runners, compiled, run and photographed inside AuToken. The sentences that produced them are the quotes.
A stadium, late afternoon
“Build me a proper tennis court in there, stadium, crowd, net and all, late afternoon.”
About 120 objects carrying 3.1 million triangles: 23,137 seats and 19,481 seated spectators in a two-tier bowl, each one generated with its own build, clothes, hair and pose, then merged by colour into a few dozen meshes. The court is at real dimensions: 23.77 by 10.97 m, with the net 0.914 m at the centre and 1.07 m at the posts.
Hong Kong, at night
“Give me Hong Kong at night, looking across the harbour from the Kowloon side.”
About 580 objects: 350 generated towers, and Two IFC, the Bank of China Tower, Central Plaza, HSBC and the Cheung Kong Centre built from primitives. The island behind is an 87,000-triangle heightfield shaped on its real summits, from Victoria Peak at 552 m to Mount Parker, with lit contour roads, beacons on the summit masts and the Peak Tower on the gap. The searchlights and the cloud on the Peak are additive light.
A cottage, with textures
“Create me a very visual 3D environment, and make sure it uses textures.”
About 380 objects. The cottage is laid out from one set of dimensions so its parts meet: 2.55 m eaves, a 45° roof in fourteen slate courses a side. A gravel drive runs through a gap in the wall to a dropped-kerb crossover; the car is a 1960s estate assembled from primitives. Stone, slate, oak, cobbles and brick are seamless textures AuToken generates at load and projects in world space.
Each took a handful of iterations: the loop saved a frame, looked at it, and rewrote the program. All three use generated materials, real-time shadows and tone mapping. Click any frame for the full 1920 by 1080 render.
What it is
Most assistants answer. AuToken executes. You give it a goal, typed or spoken, and a model works it in one session: it looks up the API, writes Java, has AuToken compile and run it inside the live app, looks at what came out, and keeps going until a judge agrees it is done or it can tell you exactly why not.
Every goal leaves a full record: the system prompt, the tools it was offered, the whole conversation, and a trace of every call with its output. The Goal Inspector reads it while the goal runs or days later. When a goal produces something visual, the evidence is the frame it rendered and what the judge made of it.
It has been my working environment every day since August 2025. Dictation, image, sound and music generation, a live 3D world with its own renderer, phone control, printer telemetry and an autonomous rover all hang off the same goal loop.
Most of it I drive from the mouse. Holding the middle button dictates into whatever app has focus, and holding right and middle together takes a spoken goal. There are 126 bindable actions in all. Input comes through Windows Raw Input rather than a low-level hook, after a hook froze the cursor twice.
It is not a model. The language, image, video and voice models it calls are rented from OpenAI, Anthropic, Google and the others, on my keys, and anything a provider painted or spoke is theirs, not mine. What is mine is everything that decides what to ask for, runs the program the model writes, checks the result and keeps the record: the application, the runner API, the agent loop and its judge, the OpenGL world, the speech stack, the phone and printer bridges, the rover and its console, the installer. That is the 516,000 lines, and every screenshot on this page is of that.
It is not a chat window with plugins. It is not a wrapper around Codex or Claude Code; it can drive them as sub-agents when a goal is repository work. And it is not a product with a team behind it. One person designs it and one person runs it, every day.
Why it exists
Nothing that ran on your own machine could do it yet. I did not set out to beat anyone to it. I wanted to find out how far one person could take the idea on their own hardware, and I have been independent since, alongside the trading platform, the research platform and the robot.
AuToken is what the plan turned into, and its repository dates from August 2025. It was never a demo. I use it for everything, and the design follows from that. Typed Java, because the compiler catches a wrong call before anything runs. A judge that is never told the request, because a model marking its own work once passed a seven-vertex pyramid as a rigged dragon. A dollar budget on every goal, because I pay the bill.
I have been doing this since 2016. The trading platform I have run alone since then already had an orchestrator that retrains its models every night, gates the result on a full test suite and redeploys itself without a human in the loop. AuToken takes the same instinct, that the machine should do the whole job and account for it, and points it at the entire desktop.
How a goal runs
Each goal is one tool-calling session with a model. It writes Java, AuToken compiles and runs it inside the live application, and a separate judge decides when it is done.
The model writes ordinary Java against sixty-three Runner classes, and run_java compiles it in-process with every runner already imported. Objects it creates stay alive between calls, so it is working in a live session inside the running app. A wrong call comes back as a compiler error the model can read.
The runner contracts used to be pushed into every turn, about 455,000 characters of them. Now the model asks for what it needs through api_overview, api_search and api_read, over about 1,800 documented methods, many with worked examples.
When the model says it is done, AuToken renders up to three views of the result. A vision model describes them without being told the goal, and a judge compares that description with the goal and scores it. An objection sends the model back to work.
Each goal carries a dollar budget, and the model is warned when three quarters of it is gone. When the conversation passes 250,000 characters or nine images, the model writes itself a handover note and a fresh episode carries on from it. For its first year a deterministic pipeline of mine repaired the model's Java before asking it again, and in September I deleted about 100,000 lines of it, because tool-calling models had made it unnecessary.
What it does
The 3D world is one corner of it. Every capability is a set of typed runner methods, and the model reaches all of them the same way: by writing Java against them.
Two lines from a real snippet, as the model wrote them, placing a wall in the 3D world and saving a frame of it:
prim.cuboid("quay_wall", "Harbour wall", I)
.sizeX(150.0).sizeY(5.0).sizeZ(30.0).center(0.0, 1.8, -71.0).color(STONE_DARK).build();
String view = cam.saveThreeDView(OUT, "oblique", -66.0, 14.0, 34.0,
24.0, 8.0, -72.0, null, null, null, 1400, false, "#3a6795");The fourteen-argument call is real. The model gets it wrong often, and the compiler tells it so before anything runs.
Windows, clipboard, mouse and keyboard input, processes, files, scheduled jobs. It can operate the machine it is running on.
Drives Chrome over the DevTools protocol with its own client and deterministic DOM and JavaScript commands, makes HTTP calls, runs SQL, reshapes JSON.
Push-to-talk and continuous dictation with a measured stop rule, every capture spooled to disk until its words are delivered, and spoken replies in configurable voices. Used more than the keyboard.
A MIDI song and track builder that knows keys and progressions, an arrangement of WAV and MIDI clips with volume and pan automation, and mix-down to WAV and MP3. Asked in one sentence for five minutes of prog rock with vocals, DeepSeek took 33 turns and came back with 5:03 in E minor through 7/8 and 5/4 on eight lanes, having written its own formant vocal synth on the way, for 25 cents.
Generation, a fluent edit pipeline, provider-backed AI edits, webcam capture, timelines, video from text or stills, screenshots and OCR, a catalogue that reuses media before spending a call.
My own OpenGL renderer, on LWJGL, inside the media panel: primitives, terrain, roads, water, trees, grass, fire, fog, skyboxes, Gaussian splats, voxels, cameras, motion paths and video capture, all driven from the model's Java, all verifiable from rendered frames. Thirty-two runners for this alone.
An Android phone through AuToken's own Kotlin app over a WebSocket, with forty-nine methods: screen share, both cameras, taps and typing, recording. ADB only installs the app. Bambu Lab printers over MQTT on the LAN, with a watchdog that sounds the phone's alarm. OpenVR.
Drive it, look through its cameras, read its telemetry and talk through it. Goals travel both ways: the desktop sends them to the rover, and the rover sends them back.
A world bible, a cast with fixed looks, a branch tree, every scene and its image prompt, built into an HTML book you click through.
Things it made, and who did what
Three things that came out of a single sentence each. In every one, a provider painted or filmed the frames, and AuToken did everything else: decided what to ask for, wrote and ran the Java that asked, kept the result consistent, checked it, and filed it.
Goal: a playable interactive novel about a cat that walks the old roads. The story runner wrote the world bible, a cast with fixed visual anchors, the branch tree, every scene and every image prompt, and built the HTML you click through. Thirty-eight frames, one cat, recognisably the same animal in all of them.
Goal: a satellite image of the whole city, and then the same streets as a medieval chart. The satellite and street-map runners fetched 4,867 tiles, projected and stitched them into one 238-megapixel image, then handed the real street graph to an image model with instructions to draw it as parchment. The roads on the chart are Cork's roads.
Goal: an ornate chessboard in a palace garden, then a six-second orbit around it. AuToken asked an image model for the board, passed that frame to a video model as the reference, pulled the first and last frames back out, and filed the still, the clip and the frames in its media catalogue so the next goal could reuse them instead of paying again.
The rover
A four-wheel robot I built from the chassis up: Teensy firmware for the encoders, a six-sensor time-of-flight ring and an IMU; ninety thousand lines of Java on a Raspberry Pi 5 for control and odometry; two Kotlin apps; a console that runs on the rover's own screen, on the desktop and on a phone. Goals reach it through the same REST surface as everything else, and it answers by voice.
Under it: a closed-form skid-steer pivot condition derived from the friction model, a measured feed-forward inverse-map wheel controller in place of PID, fused encoder and IMU odometry, and an overhead computer-vision ground-truth lab (OpenCV, ChArUco) accurate to about three millimetres on the floor, used to calibrate the estimate rather than paper over it with gains.
Integrations
Fifteen provider hubs sit behind the badges, most of them over one shared base. Every model it offers is listed with its context window, the reasoning levels it actually accepts, and its price per call: 61 language models from six vendors, routed across fifteen jobs.
Codex and Claude Code do repository work better than I would by hand. AuToken can hand them the repository work on my own subscriptions, sharing its MCP tools with them and keeping the final check for itself, and carry on with everything they cannot do: talk, draw, render, drive a phone or a robot. When one subscription runs out it tries the other, then its own paid loop.
Versus the coding agents
Codex and Claude Code edit repositories from a terminal, and AuToken uses them for exactly that. What it does itself is the part they were never built for.
| Aspect | Terminal coding agents | AuToken |
|---|---|---|
| What the model produces | Edits to files, shell commands | Java compiled and run inside a live desktop app, against 63 runners and about 1,800 documented methods |
| When a wrong call is caught | At runtime, when the command fails | At compile time, and the compiler's message goes straight back to the model |
| Checking the result | Tests pass, or the model says so | A vision model describes the render without being told the goal, and a judge scores that against the goal and can refuse "done" |
| The record | The session transcript | System prompt, tools, conversation and a trace of every call for each goal, open in the Goal Inspector |
| Input | Typed prompt | Typed, spoken, dictated into any app, or read from the clipboard and images |
| Output | Text and files | Speech, images, video, sound, music, a rendered 3D world, files, actions on devices |
| Reach | The working directory | The desktop, Chrome, a phone, printers, a headset, a robot |
| Models | One vendor for Codex and Claude Code; Cursor and Copilot offer several | Six language vendors and more than ten media vendors, with a dollar budget on every goal; runs on your keys |
| Where it runs | A terminal session | A desktop application with a local REST and WebSocket API, so scripts can do anything the window can |
| Unattended | Until the session ends | Alarms, scheduled goals, and an away mode that keeps working |
| Uses the coding agents | n/a | Yes: Codex CLI and Claude Code are sub-agents it can hand goals to, on my subscriptions, with its own final check |
| Limits | Polished products with teams behind them, far stronger at repository work: indexes, diff review, PR agents | No sandbox: the model's Java runs in-process with full access. Windows only, one user, thin tests, and a renderer short of PBR |
Engineering
I designed AuToken and directed every part of it, and coding agents wrote much of the code.
| Java | 516,000 lines across 1,802 files in five Maven modules, down from a peak of 601,000 before the September clear-out |
|---|---|
| Also | Kotlin for two Android apps, Python and Teensy firmware for the rover, JavaScript for the web surface |
| Commits | 22,840 between 12 August 2025 and 24 September 2026, on 236 active days. About 9,100 of them merge Codex cloud tasks, and 565 carry a Claude co-author line |
| Runners | 63 exposed to the model with about 1,064 methods; ThreeDRunner alone exposes 194 |
| Provider hubs | 15 hubs, 11 over one shared base, with model lists and cost tracking; 61 language models from six vendors |
| Agent loop | One tool-calling session per goal over 24 tools, with a blind judge, a dollar budget and a handover when the context fills |
| Tests | 590 JUnit tests, thin for the size of it; a material check on every 3D build and the blind judge on visual goals |
| API | About 49 REST endpoints and a WebSocket on 127.0.0.1, and an MCP endpoint serving 21 of the agent's tools |
| Packaging | A Windows installer with its own runtime, FFmpeg and VLC; a first-run window that verifies keys with a real call |
| Demos | Sixteen double-clickable tutorials in six groups that reset the workspace and run unattended |
A screened desktop build exists as a separate private repository, with the runner catalogue, the REST surface and a README written for a stranger.
Sixteen double-clickable tutorials ship with it. Each resets the workspace and runs unattended while a voice explains what is on screen.
Everything below the provider line is mine. Everything above it is rented, per call, on my keys.
sportyai: e-sports prediction models that beat the bookmakers for years, ~220,000 lines, run alone. Its orchestrator retrains models nightly, gates on the full test suite, redeploys itself.
Left salaried work, the same weeks Auto-GPT and BabyAGI first let a model execute its own code. Independent since.
crypto-ai: a quantitative research platform across fifteen exchanges, live level-two books on eight of them.
The AuToken repository, and within weeks the 3D world with its own renderer. From the first month it was the environment the rest of the work happened in.
The rover, phone and printer control, the installer, the screened public build.
Goals moved to one tool-calling agent with a blind judge. 22,840 commits and counting.
Contact
Fifteen years of Java on real-time trading systems, then three of this. Available immediately, remote, based in Cork.