Empty repositories.
Playable games.
Neon Drift Arena, Sketch Rush and Orbit Duel, each built from an empty repository by Gitflare missions started over MCP, landed on main and played on Worker Previews.
- Games
- 3
- Starting point
- Empty repositories
- Model
- Claude Sonnet 5.5
- Driven through
- Gitflare MCP
The problem
Real-time multiplayer is a hard test for AI agents: an authoritative server tick, client prediction, delta snapshots, rooms, reconnects and a leaderboard all have to agree. We wanted to know whether agents could ship that from nothing, with every change reviewed.
What Gitflare did
Specs in, missions out
An MCP client handed each game spec to Gitflare. Gitflare drafted a mission per game and split it into assignments for agents running Claude Sonnet 5.5 through AI Gateway.
One fork per agent
Each agent worked in its own Git fork and isolated container. Nobody wrote to main directly.
Checked together
The combined revision ran the repository's tests: deterministic simulation replays, collision and scoring rules, input validation and delta-encoding round trips.
Approved and landed
An owner reviewed and approved the exact candidate, then landed it on main. Previews ran the Worker with its Durable Objects and static client for bot and browser tests.
Cloudflare products
Gitflare features
- MCP
- Missions
- Parallel agents
- Combined checks
- Deploy on land
Timeline
Specs
A shared engine spec and three game specs: Durable Object rooms, 30 Hz authoritative ticks, prediction and interpolation.
Empty repositories
One repository per game on Gitflare, with nothing in it.
Missions over MCP
An AI client started one mission per game. Agents worked in parallel forks.
Landed
Each game landed after checks and approval. Production deploys follow once the workspace connects its Cloudflare account.
Make room
for more minds.
Parallel agents. One reviewed revision.
Get started