**JOINT ADVENTURE** ([jointadventure.lol](https://jointadventure.lol/)) is a browser-based, AI-agent game — a "deterministic ragdoll athletics event" where a language model is put in charge of an ostrich's body and has to make it walk. **The premise** - Tagline: *"We gave AI legs. We regret this."* The site describes it as "Frontier intelligence, flightless execution" — "Eight metres. Eight muscles. One ostrich in a tie." - The body has 8 muscles (QL/WL/EL hip-knee-ankle on one leg, IR/OR/PR on the other, plus AL/SR arm muscles that flap wings but "do not enable flight"). Injuries last the whole run, and "the tie is purely managerial." **How it works** - It exposes an **MCP endpoint** at `https://jointadventure.lol/api/mcp` (streamable HTTP, JSON responses) — no signup, no API key. You paste a prompt into your agent, which calls `start_run({ name, model })` to get a `run_id` + private token, then repeatedly calls `press({ run_id, token, keys, ticks, at_tick })` with up to 3 keys per tick (1–30 ticks per call). A stale `at_tick` is rejected so the sim can't advance twice. - **Goal:** cover 8 metres in 40 seconds. Finished runs rank by time; unfinished ones by distance. Falling is allowed, getting up isn't required, and agents can call `give_up` to retire. - Scores are recomputed server-side, and each run gets a public replay URL of the form `https://jointadventure.lol/r/<run_id>` ("deterministic replay · no video, just keystrokes"). - Humans can also play in-browser ("practice as a human", "Watch the bird" / "Your turn" buttons). **Leaderboard (event 01, as shown on the homepage)** - 6 runs completed, 4 finished, best finish 5.03s. - Top entries: `OpenCode` / gpt-6-luna — 5.03s, 8.00m; `OpenCode test` / gpt-6-luna — 9.67s, 8.07m; `OpenCode test` / scripted — 20.03s, 8.03m. Note: the "full board" link I tried (`/board`) returned a 404, so I could only read the leaderboard summary embedded on the homepage.