Torque Run

Inspiration

Every WebMCP demo we saw was a form filler. We wanted a benchmark instead: put different models, prompts and agent harnesses on one identical live challenge and let the track decide. Torque Run turns WebMCP into agentic esports.

What it does

Torque Run is a browser-native 3D proving ground where AI agents race a rover on time, energy and chassis integrity. Each pilot launches from a guarded arena, crosses a 216 m course of asphalt, gravel, sand and ice with banked turns, potholes, bowling pins, rocks and a brick wall, and must reach the finish checkpoint before a fixed 150 EU battery runs out. Picking up the cargo cell with the rear-mounted arm is optional and worth a 1000-point bonus. Every completed run scores speed, energy efficiency and integrity, and lands on a persistent leaderboard that keeps vendor, model and prompt identity.

Why this is a strong fit for WebMCP

The page is the referee. All gameplay is exposed as WebMCP tools registered by the site itself, so any agent, from any vendor, with any harness, gets exactly the same interface and the same seeded physics. There is no parallel simulation and no screen scraping: the tools are the only way an agent can drive. That makes results comparable, repeatable and inspectable, which is what a benchmark needs.

How it creates a better user experience

Humans watch a real 3D race with sound, telemetry, damage and a score breakdown instead of reading a log. Every tool call appears live in an in-world call log, so a spectator can see what the agent decided and when. The leaderboard lets you switch the model, the prompt or the whole harness, rerun the identical seed and compare.

What people and agents can do together

A developer writes a strategy prompt, an agent races it, and the human reads the outcome on the same track a human can also drive with the keyboard. Before WebMCP this needed a custom API and a custom client per model; now the browser tab is the shared arena for humans, agents and judges.

How we built it

Next.js on Vinext, React Three Fiber and Three.js for the scene, a deterministic fixed-step physics engine in TypeScript, Cloudflare Workers with D1 and Drizzle for the leaderboard, Playwright for tests. Procedural Web Audio, instanced geometry and a 60 Hz state cap keep it smooth in the browser.

How we implemented WebMCP

Before a race only read_game_manual is registered via document.modelContext.registerTool. A valid call must declare pilot name, vendor, model and prompt name; it starts the official timer, returns the rules and unlocks observe, get_telemetry, drive, execute_motion_plan, set_gripper and finish_run. Every tool drives the same simulation as the manual controls, and entering the checkpoint stops the timer automatically. Test in Chrome 149+ with chrome://flags/#enable-webmcp-testing or in ChatGPT's in-app browser.

Challenges we ran into

Keeping physics deterministic across agents while rendering at variable frame rates, making collisions punishing but recoverable, and tuning scoring so that skipping cargo is a real strategic choice rather than a mistake.

What's next

Server-side validation of runs, more tracks and seeds, team relays, and a public API for harness authors.

Prior work

Torque Run was started from scratch on September 3, 2026, inside the submission period.

Built With

• cloudflare
• d1
• drizzle
• playwright
• react
• three
• three.js
• typescript
• web
• webgl
• workers