darrenhinde/OpenAgentsControl

What it solves

Most AI coding assistants generate generic code that doesn't match your project's conventions, forcing you to spend hours refactoring and rewriting. OpenAgentsControl (OAC) fixes this by teaching AI agents your coding patterns up front, so they produce code that fits your tech stack, naming rules, validation style, and architecture from the very first attempt. It also adds human approval gates so agents never make changes without your sign-off.

How it works

OAC is built on top of the open-source OpenCode framework and extends it with specialized agents and a context system. You describe your patterns once (via the /add-context wizard or by editing markdown files), and a subagent called ContextScout discovers the relevant patterns before any code generation. Agents then propose a detailed implementation plan, wait for your approval, and execute incrementally with validation. The design follows a Minimal Viable Information (MVI) principle: context files are kept under 200 lines and loaded lazily only when needed, reducing token usage by up to 80% compared to loading the whole codebase. OAC includes multiple agents (OpenAgent for general tasks, OpenCoder for production features, SystemBuilder for custom AI systems) and subagents for testing, code review, documentation, and more. It also has an ExternalScout that fetches live documentation for external libraries to avoid outdated training data.

Who it's for

Developers and teams who already have established coding standards and want AI help that respects those standards. It's aimed at production code, team collaboration, and anyone who wants more control and predictability from AI coding assistants. It works with any language and any model (Claude, GPT, Gemini, local models) and is available both as an OpenCode agent set and as a beta Claude Code plugin.

Highlights

• Pattern learning: teach agents your code style once, get matching code every time.
• Approval gates: agents always propose a plan and wait for your approval before writing or editing files, running commands, or delegating.
• Editable agents: agent behavior is defined in markdown files you can edit directly—no compilation or vendor lock-in.
• Token efficiency: MVI design keeps context small and lazy-loads only what's needed, cutting token costs.
• Team-ready: context files can be committed to the repo, so all team members (and new hires) automatically inherit the same standards.
• Model agnostic: works with any AI model or provider, including local models.