Most AI coding tools optimize for raw generation quality; this one optimizes for the moment before generation, betting that the real cost of agentic coding is rework, not tokens. Built as a layer on top of the open-source OpenCode CLI, it inserts two gates into the loop: a ContextScout pass that distills your repo's existing conventions into a compact brief, and a human approval step where you sign off on the agent's plan before a single line is written.
What Sets It Apart
- Plan-first, approval-gated. A multi-stage workflow routes work through specialized subagents (context scout, coder, test engineer, code reviewer) but pauses for your sign-off, so a wrong assumption costs a review click instead of a full refactor.
- Context as a committed artifact. Project patterns live as markdown under
.opencode/context/and ship with the repo, so the same house style applies whether you or a teammate runs the agent. - Minimal Viable Information (MVI). Instead of stuffing whole files into context, it loads only the patterns a task needs; the author reports roughly 80% less context overhead, which matters most on large codebases.
- Editable, not black-box. Agents are plain markdown you can read and rewrite, and it runs against Claude, GPT, or Gemini rather than locking you to one provider.
Who It's For
Great fit if you have an established codebase with conventions worth enforcing and a team that wants AI output to match them by default — especially if you already use OpenCode or want the companion Claude Code plugin. Look elsewhere if you want zero-config autopilot: the approval gates and context setup are deliberate friction, and the payoff only appears once you've invested in writing your patterns down. As a solo, MIT-licensed project rather than a vendor product, expect fast iteration over long-term support guarantees.
