Inspiration

Coding agents can work in parallel, but supervising them usually means scanning five streams of logs, tool calls, diffs, and chat. Codeville asks a different question: what if five agent sessions felt like one living town?

What it does

Five local repositories become five workshops, each with its own real gpt-5.6-sol Codex thread through one local app-server connection. As a thread plans, reads, edits, tests, or completes, its builder moves through the corresponding workshop state, so concurrent work is legible at a glance.

Two things make it a command center rather than a dashboard. First, nothing an agent does touches your repository until you approve it: every session runs in a scaffold (a git worktree on its own branch), and finished work arrives as a pennant. Site inspection shows a deterministic fact line the model cannot claim, a builder's account in which file names survive only when the diffstat vouches for them, and the full diff. One click installs a single squash commit. Second, supervision is specific without turning code into spectacle: the Village feed streams every builder's real steps in one place, while prompts, diffs, and command output never cross into the visual layer. Wall mode drops to a provably code-free subset, safe to leave on a shared screen. You can also redirect a builder mid-turn, queue work orders, and answer builder questions with context, right at the desk.

How we built it

Codex with GPT-5.6 built Codeville itself: the architecture, generated app-server protocol bindings, the Electron/React/PixiJS implementation, the scaffold engine, the privacy sanitizer, migrations, tests, and packaging. Human decisions set the boundaries: native and local-first, one stdio connection instead of hosted orchestration, validated outcome debriefs instead of raw model text, and nothing merged without the user.

Challenges and what we learned

The hard parts were proving real concurrent app-server behavior (five threads multiplexed over one connection without cross-talk), designing a debrief contract the model cannot fake, a canvas-recreation glitch that would have reset the village on every update, and an automated release gate that starts five real GPT-5.6 threads, proves five checkouts stay byte-identical, installs all five improvements, and relaunches the app to prove persistence. The demo video shows the same flow running on the maintainer's five real repositories.

What's next for Codeville

Support for more agent CLIs alongside Codex, more than five lots for bigger portfolios, Windows and Linux builds, and signed releases. Deeper in: standing per-repository trust policies so routine approvals stop needing a click, and richer work-order queues so the town can run a full day's backlog while you stay the merge gate.

Built With

Share this project:

Updates