I made a more complete coding harness then most
I've been building ACE: you point it at a goal and it plans, writes, tests, reviews itself three ways, and opens PRs — mostly hands-off. It's early (v0.0.1-alpha) and I'd genuinely love blunt feedback.
🔗 github.com/buagi/ace (demo at the top)
What it does, plainly: You describe the project once — stack, containerized or not, who the users are, how it deploys — then run ace autorun. A 9-agent crew takes over: an orchestrator plans the work; an implementer writes it; a test engineer adds tests; and reviewers check logic, UX, and standards before anything merges. CI/CD is baked in; it can deploy to a VPS over SSH. You pick auto-merge or approve-each-merge.
What's actually different:
- Works on several features at once (
ace swarm) — each in its own git worktree. It uses code-graph tools (GitNexus/Serena) to understand each change's impact, so it knows what's safe to parallelize and resolves the predictable conflicts itself. - Cheap + self-hosted. The 8 worker agents run on DeepSeek V4; only the overseer uses a stronger model (Claude/GPT — or DeepSeek too for zero subscription). Installs user-local, no root — built and daily-driven on Fedora Silverblue.
- Grounded, not vibes: cross-checks structure via GitNexus/Serena and live library/version docs via Context7, so it hallucinates less.
Try it in 2 min, $0: ace loop dash --demo replays the whole loop — no keys, nothing installed. MIT licensed.
It's rough in places and I know it's early — I'd love feedback on: is the setup too heavy? does the parallel-swarm idea actually make sense to you?
It's fully MIT licence
