AI Agents That Learn Your House Style—Then Ask Before Acting

What it does

OpenAgentsControl is an extension of the OpenCode CLI that layers specialized agents, context management, and team workflows on top. It prompts you once for your tech stack, conventions, and security requirements, then loads those patterns before generating any code. Every plan requires explicit human approval before execution—no surprise file rewrites. The context files live as editable markdown in your repository, so standards propagate through git like any other config.

The interesting bit

Instead of dumping the entire codebase into the prompt, a “ContextScout” agent discovers relevant patterns lazily; the project claims this cuts token usage by roughly 80% compared to loading full context. Because the agents themselves are just markdown files you edit directly, you’re not locked into opaque, baked-in plugin behavior.

Key highlights

• Mandatory approval gates: Agents must propose a plan and get a human yes before implementing.
• Pattern inheritance: Commit .opencode/context/ to git and new team members inherit standards automatically.
• Model agnostic: Works with Claude, GPT, Gemini, MiniMax, or local models.
• Multi-language support: Targets TypeScript, Python, Go, Rust, C#, and others.
• Claude Code plugin: Also ships as a beta plugin for Claude Code with seven specialized subagents.

Caveats

• Execution is strictly sequential with approval steps, so it trades speed for control compared to fully autonomous tools.
• The Claude Code plugin is explicitly marked beta.
• Requires the OpenCode CLI as a prerequisite, so it is not a standalone replacement for your existing setup.

Verdict

Best for teams with established conventions who want reproducible, reviewable AI assistance rather than a fast autonomous coder. Skip it if you prioritize raw speed over governance or need a completely self-contained tool.