codex-collab
by Kevin7Qicodex-collab is a Claude Code skill for handing work to Codex through the codex app-server JSON-RPC protocol. Use the codex-collab skill to run tasks end-to-end, review code, research in parallel, and support workflow automation with resumable threads and reusable output.
This skill scores 78/100, which means it is a solid listing candidate for directory users: it has a real, actionable workflow for delegating work to Codex, but users should still expect a few adoption caveats around setup and usage details.
- Explicit trigger language for invoking, delegating to, or collaborating with Codex, which helps agents decide when to use it.
- Substantial operational detail: run, review, parallel work, and research modes are described, with JSON-RPC protocol framing and example commands.
- Repository evidence shows a real CLI implementation with tests, install scripts, and CI, which increases trust that the workflow is executable rather than placeholder-only.
- No install command is embedded directly in SKILL.md, so users may need to rely on repository scripts or README/install docs to set it up.
- The skill depends on external prerequisites like Bun and Codex CLI/app-server support, which narrows adoption to users with that environment.
Overview of codex-collab skill
codex-collab is a Claude Code skill for handing work to Codex through the codex app-server JSON-RPC protocol. Use the codex-collab skill when you want a second model to run a task end-to-end, review code like a PR reviewer, or investigate something in parallel while you keep working.
What this skill is for
The codex-collab skill is best for tasks where a direct Codex run is more useful than a generic prompt: implementation, analysis, research, and review. It turns a vague “ask Codex about this” idea into a structured workflow with threads, resumable turns, and output you can reuse.
Why people install codex-collab
Most users care about speed, repeatability, and not losing context. codex-collab install gives you a CLI-backed bridge instead of a one-off prompt, so you can run a task, resume it, review output, and keep the thread identity across steps. That matters when you are testing changes, comparing options, or delegating a narrow job to Codex.
Best fit and misfit cases
Install codex-collab if you already use Claude Code, have the Codex CLI available, and want Codex to participate in your workflow as a separate agent. It is a strong fit for codebases, architecture questions, and workflow automation. It is a weaker fit if you only want a single static answer, do not have Bun or Codex installed, or do not need thread-based collaboration.
How to Use codex-collab skill
Install the skill and verify prerequisites
For codex-collab install, you need Bun and the Codex CLI with app-server support. The repo provides install scripts for Unix and PowerShell, and the default install is self-contained while --dev keeps symlinks for live iteration. After installing, run a health check or a small read-only task before trusting it on real work.
Start with the right input shape
The most useful codex-collab usage starts with a concrete task, not just “help me with this repo.” Include the goal, repository or working directory, constraints, and the output you want from Codex. Good inputs name the deliverable: patch, review, plan, investigation, or explanation. For example: “Review src/cli.ts for unsafe argument handling and return only actionable findings.”
Suggested workflow and files to read first
Read SKILL.md first, then README.md, CLAUDE.md, package.json, and src/cli.ts to understand commands, defaults, and how threads are managed. If you are installing or debugging, inspect install.sh, install.ps1, and src/protocol.ts before changing anything. For workflow automation, the most important files are src/threads.ts, src/turns.ts, src/events.ts, and src/approvals.ts, because they show how jobs start, resume, emit events, and request approval.
Prompt patterns that work well
Use one clear command per job. codex-collab run is the best default for prompted tasks; review is better for code review; multiple jobs are useful when you want parallel research. Mention whether Codex should be read-only, implementation-focused, or review-focused, and specify whether it should continue from an existing thread. That level of structure materially improves codex-collab usage for workflow automation.
codex-collab skill FAQ
Is codex-collab only for code?
No. The codex-collab skill is designed for code, but it also fits research, planning, and repository analysis when you want a separate Codex perspective. The practical limit is whether Codex can act on the task through its supported app-server workflow.
Do I need this instead of a normal prompt?
If you just need one answer, a normal prompt may be simpler. Use codex-collab when you need a runnable task, resumable context, or structured review output that can be repeated and audited. That is the main reason to choose codex-collab guide style workflow over ad hoc prompting.
Is it beginner-friendly?
Yes, if you follow the repo’s existing command flow and keep the first task small. The main friction is setup, not concept: Bun, Codex CLI, and the app-server protocol must all be working. Start with a read-only research task before asking for edits.
When should I not use it?
Do not use codex-collab for Workflow Automation when you need a pure local script with no Codex dependency, or when your environment cannot support Bun or the Codex app server. It is also overkill for tiny questions that do not benefit from thread state or delegated execution.
How to Improve codex-collab skill
Give Codex the missing context up front
The strongest codex-collab usage includes the target files, the acceptance criteria, and the constraints that matter. Say what should not change, what a successful result looks like, and whether the output should be a summary, patch, or review. This reduces vague outputs and avoids follow-up turns just to recover basic context.
Use the right mode for the job
If you want an answer, use run. If you want a second set of eyes, use review. If you are exploring, use read-only. Choosing the mode correctly is one of the biggest quality levers in codex-collab, especially when automating multi-step work.
Watch for common failure modes
The main failure mode is under-specifying the task, which leads to generic or overbroad output. The second is mixing incompatible goals, like asking for both a review and a rewrite in the same turn. A third is skipping repo-specific constraints, such as preferred runtime, sandboxing, or approval behavior.
Iterate with thread-aware follow-ups
After the first result, refine the prompt with concrete deltas: “focus on src/protocol.ts only,” “now compare against CLAUDE.md,” or “resume the same thread and test the edge case.” That is where codex-collab becomes more valuable than a one-shot prompt, because it preserves context and makes improvement incremental.