# Claudexor Claudexor by Anton Razzhigaev is a local-first control plane that sits ABOVE the AI coding agents. One interface orchestrates [[Claude Code]], Codex, Cursor, and OpenCode: coordinate them, race them against each other, review one model's work with another, and audit everything, without committing to a single vendor. Read-only: `ask`, `plan` (with `--council` running parallel plans across harnesses and merging them), and `deep-scan` for bounded multi-agent research. Mutating: `agent` for normal execution and `best-of`, which races multiple candidates, reviews them, and picks a winner. On top sit cross-model review with typed approvers, deterministic test gates, protected-path approvals, and hard budget caps that reject runs with unknown costs. Architecture: - **Write turns run in isolated git worktrees.** Nothing touches your repo until you approve. Read-only turns resume the vendor's native session, so thread continuity survives. - **Files are the source of truth.** Every run leaves inspectable artifacts on disk; the terminal UI is just a projection. Every claim (cost, quota, auth, web evidence) is a typed fact you can check. No telemetry. - **Multi-profile credential management** with per-profile quota tracking and opt-in rotation when a subscription hits its limits. MIT licensed, with CLI surface. This is the most concrete implementation I've seen of two ideas at once: [[Harness Engineering]] pushed up a level (a harness OF harnesses), and something close to the [[Living shim between humans and AI models]]: a layer that manages the models so you manage outcomes. ## References - Official website: https://claudexor.ai/ - GitHub: https://github.com/razzant/claudexor ## Related - [[Claude Code]] - [[AI Agent Harness]] - [[Harness Engineering]] - [[Living shim between humans and AI models]] - [[Orchestrator tax]] - [[AI Agents]]