The latest entry in the agentic coding ecosystem is claude-codex-pair, a lightweight tool that orchestrates a peer-review loop between Anthropic’s Claude Code and OpenAI’s Codex. Rather than relying on complex API integrations or proprietary middleware, the tool facilitates communication through a shared directory on the local filesystem, allowing two live terminal sessions to exchange requests, reviews, and responses without leaving the developer’s machine.

Disk-Based Protocol Over API Complexity

The core mechanism is surprisingly low-tech: agents write Markdown files to ~/.pair/-//, creating a persistent audit trail of the interaction. The workflow is symmetric, meaning either agent can act as the author or reviewer. When one agent finishes a task, it sends a request via pair send, and the other agent polls for updates using pair wait. This approach bypasses the latency and cost concerns of continuous API polling, grounding the multi-agent interaction in standard filesystem operations.

Strict Sandboxing and Configuration

Implementation requires specific permissions, as both agents’ sandboxes must have write access to the shared channel directory. For Codex, this involves updating ~/.codex/config.toml to include writable_roots, while Claude Code relies on a generated skill file in .claude/skills/pair/SKILL.md. The tool also enforces a hard limit on review cycles; if findings remain unresolved after three rounds (configurable via PAIR_MAX_ROUNDS), the thread escalates, forcing both agents to halt and request human intervention.

Git Worktree Isolation

A notable feature for power users is the native support for Git worktrees. Each worktree generates a unique channel hash, allowing developers to run parallel pairing sessions on different branches without message cross-contamination. The pair status command provides real-time visibility into the active lanes and verifies that the instruction files on disk match the running agent’s configuration, reducing the friction of debugging stale agent states.

Key Takeaways

  • The tool uses a simple Markdown-over-filesystem protocol to avoid complex API dependencies.
  • Review cycles are capped at three rounds by default to prevent infinite agent loops.
  • Native Git worktree support enables parallel development streams with isolated agent contexts.

The Bottom Line

This is a pragmatic step toward agentic self-correction that prioritizes transparency and local control over cloud-based orchestration, though its reliance on manual sandbox configuration may limit immediate adoption.