VOLY — Control Plane for AI Agents
A self-hosted control plane above coding agents such as Claude Code, Cursor and OpenCode that routes tasks between them, tracks and limits spend, and runs multi-agent workflows.
VOLY sits above coding-agent "executors" such as Claude Code, Cursor, DeepSeek and OpenCode, decomposing work into role-based sub-agents, such as architect, developer, tester, reviewer and devops, with per-role model tiers, and can run some roles through executors while keeping others as chat. It guards file writes with dry-run diff previews, protected paths, a soft rollback and a git-based rollback, controls spend through Cloudflare AI Gateway and configurable limits, and reduces token use with a persistent cache and model routing.
It can search GitHub for reusable code before generating new code, pins a project's tech-stack versions before a run, and offers greenfield scaffolding for empty projects. Multi-agent steps are checked against plan gates with scoped test runs, and outcomes are evaluated with deterministic policies, golden-regression replay and optional rubric-based LLM judges. It supports DSPy as an optional optimization layer and stays project-agnostic through a configurable target directory.