Dev & Engineering codex-cli-rpccode-reviewthread-managementpeer-messagingtool-approvals

Codex-Collab: Collaborate with Codex from Claude Code

A Claude Code skill that lets you dispatch tasks, code reviews, and parallel research to Codex without ever leaving your Claude session.

FollowSkills review · FSRS-2.0
Use with care
56/ 100 5-point scale 2.8 / 5
1 2 3 4 5 6
1Trust13 / 25 · 2.6/5

SKILL.md discloses data flow and state directory (~/.codex-collab), offers a read-only sandbox, configurable approval policies, confirmed updates, and permission-class gating of peer replies — good transparency; but it mandates dangerouslyDisableSandbox=true for ALL commands, defaults to --approval never (auto-approving every action), and exposes a danger-full-access path, so least privilege and forced confirmation are incomplete — deducted.

2Reliability10 / 20 · 2.5/5

Docs are self-consistent with exit-code tables, health checks, degradation paths, and designed failure feedback; the repo declares tests/typecheck and CI smoke-tests install+health. Static review cannot execute; test files themselves not provided, so capped at 10, and failure-feedback quality on abnormal input is only asserted.

3Adaptability9 / 15 · 3.0/5

Trigger conditions are clear (user asks for Codex collaboration or an independent perspective), with documented fallbacks and a Chinese README; but core function depends entirely on the Codex CLI and OpenAI services — a mainland-China reachability concern — and non-fit boundaries are thinly described, so deducted.

4Convention13 / 15 · 4.3/5

MIT license, version 0.3.2, release/update path, installer, contribution docs, bilingual README are all present; SKILL.md is well layered with a CLI reference. Publisher identity unverified, no standalone changelog file, maintenance rests on a single author — not full marks.

5Effectiveness6 / 15 · 2.0/5

Target scenarios (PR review, long-task delegation, thread resume) are clear and CI smoke tests support basic installability; but static review cannot verify outputs are directly usable, and comparative benefit over the official Codex plugin/MCP is author-claimed only — capped at 6.

6Verifiability5 / 10 · 2.5/5

A real CI workflow (three-platform smoke tests), declared contract tests, and bilingual docs provide auditable primary material with cross-corroboration; but the test suite content and execution results are not in the provided files and key claims (Guardian, peer messaging) cannot be independently reproduced — static cap of 5.

Evidence confidence:Low Reviewed Sep 10, 2026 Reviewed revision 47a73a7dc62c
Before you use it
  • SKILL.md mandates disabling the sandbox (dangerouslyDisableSandbox=true) for all codex-collab commands and defaults to --approval never (auto-approving everything); use -s read-only or --approval on-request/untrusted in sensitive projects.
  • Core function depends on the OpenAI Codex CLI and its services, which may be unreachable from mainland-China networks; peer messaging additionally requires Claude Code >= 2.1.224.
  • The installer writes outside the repo (~/.claude/skills, ~/.local/bin) and updates pull from GitHub releases with local rebuilds — review diffs before confirming.
  • Publisher identity is unverified by FollowSkills; this score is a static source review with no execution or independent testing.
See the full review method →

What does this skill do, and when should you use it?

codex-collab bridges Claude and Codex over Codex's app-server JSON-RPC protocol, giving you structured, event-driven task execution, code review, and thread management. It supports thread resumption, configurable tool-call approval policies, and a two-way ask channel where Codex can pose questions mid-task. On messaging-capable Claude Code installs, Codex registers as a native peer that Claude sessions message directly. The project requires Bun and the Codex CLI, ships under the MIT license, and includes unusually thorough documentation plus a self-update mechanism.

Runs prompted tasks via codex-collab run (with --resume to continue threads, --detach for background runs); performs structured code reviews via codex-collab review in a read-only sandbox against PRs, uncommitted changes, or specific commits; inspects thread state and history through threads, progress, output, and peek; handles Codex's mid-task questions through ask/answer/next; manages tool-call approvals via approve/decline, including Guardian autonomous review; and optionally starts a workspace broker with peer up that registers Codex as a cross-session messaging peer in Claude.

  1. A Claude Code user who wants an independent second opinion from Codex on code, plans, architecture, or design decisions
  2. A developer getting a one-command structured review from Codex in a read-only sandbox before opening a PR
  3. A long-term project maintainer reusing Codex thread context to chain related tasks, such as survey call sites then migrate them
  4. Someone running a large refactor detached in the background while watching live progress in a separate pane with follow --watch
  5. An automation builder running multiple concurrent Codex threads and answering Codex's mid-run questions without polling

What are this skill's strengths and limitations?

Pros
  • Structured JSON-RPC communication with typed, parseable events and real-time progress streaming
  • Thread reuse and resumption (--resume) preserves conversation context across follow-ups
  • Fine-grained approval control including the Guardian auto-reviewer with overridable denials
  • Fail-open two-way ask channel — an unanswered question never stalls a run
  • Exhaustive README and SKILL.md with exit-code tables, error-recovery tables, and self-update machinery
Limitations
  • Hard dependency on two external runtimes: Bun and the Codex CLI, both required
  • Native peer messaging limited to macOS/Linux with Claude Code >= 2.1.224; Windows falls back to CLI mode
  • Reviews are one-shot — you cannot follow up on the same review, only fork via run --resume
  • Defaults to auto-approving all Codex actions (--approval never), so high-stakes setups must tighten this manually
  • Tested on Ubuntu 22.04, macOS, and Windows 10 per the README, but the source gives no independent evidence of test-suite coverage

How do you install this skill?

Prerequisites: Bun >= 1.0 and the Codex CLI (npm install -g @openai/codex); Linux, macOS, and Windows 10 are supported. Steps: 1) git clone https://github.com/Kevin7Qi/codex-collab.git && cd codex-collab; 2) run ./install.sh on Linux/macOS or powershell -ExecutionPolicy Bypass -File install.ps1 on Windows; 3) verify with codex-collab health. The skill installs to ~/.claude/skills/codex-collab/ and is discovered automatically by Claude Code; if this is your first skill, restart Claude Code once.

How do you use this skill?

After installation, Claude invokes the skill automatically per the SKILL.md trigger conditions. Manual examples: codex-collab run "what does this project do?" -s read-only --content-only for a research task; codex-collab review --content-only for a PR-style review of the default branch; codex-collab run --resume <id> "now check error handling" to continue a thread. Use --detach for long tasks and codex-collab follow --watch in another terminal to observe. Note: run/review should run in the background and notify you on completion — do not poll.

How does this skill compare with similar options?

The README names two official alternatives: OpenAI's Codex MCP server (better for simpler interactions) and the official Codex plugin for Claude Code (built around slash commands you invoke yourself). codex-collab's differentiator is natural-language-driven use — thread management and background execution are handled for you rather than requiring per-command invocations.

FAQ

Does this skill incur extra API costs?
It communicates through your locally installed, authenticated Codex CLI, so Codex-side model usage costs apply; the source does not specify billing details — it depends on your existing Codex authentication and subscription.
Can Codex modify my files freely?
The default sandbox is workspace-write; use `-s read-only` to restrict it. The default approval policy auto-approves everything (--approval never); tighten with --approval on-request/on-failure/untrusted or --approval auto (Guardian decides autonomously).
What happens if a task times out or crashes?
Default per-turn timeout is 1 hour — a timeout usually means a stuck process, not a slow one. Thread state persists, so resume with `--resume <id>`. Exit code 5 means it died blocked on an approval; resume with a longer --timeout or --approval auto.
Will it work without a peer-messaging-capable Claude Code version?
Yes. Set `config mode cli`, or run on Windows, and everything routes through the CLI path with identical functionality — you just lose cross-session messaging and the consult tool.

Related skills