Connect

Free, open-source Claude Code plugin

Let Claude Code get a second opinion from Codex.

Connect plugs OpenAI Codex into Claude Code. When a change is risky, Claude sends it to Codex for review, checks what Codex found against your code, and fixes what holds up — all in the same session.

Install in 2 commands See a real session ↓

A real session

Claude rewrote a README, then had Codex review it.

The user asked Claude Code to rewrite this project's README and get GPT's advice. Claude wrote it, called Codex through Connect, verified the findings, and fixed them. Trimmed for length.

● ● ●claude — ~/code/Connect
❯ read this dir and write a readme.MD and ask for gpt's advice
⏺ There's already a README.md, so I'll read it and the source before changing anything.
Read 1 file, ran 4 shell commands
⏺ Write(README.md)
⎿ Added 196 lines, removed 131 lines
⏺ Next I'll ask Codex (GPT) to review it.
Connect
⏺ connect - codex_ask (MCP)(intent: "code_review", prompt: "Review README.md against the source. Which claims are wrong?")

Codex replied:

1. Lease protection is overstated — leases live in one MCP process's memory (mcp/connect-mcp.mjs:57).

2. Sharing threads with the Codex TUI only works if the TUI is started with codex --remote.

3. codex_history is not a full replay — 20 turns, 400 chars per item.

4. The README never says approval requests from Codex are auto-declined.

thread: 019a… · routed: code_review · effort high · sandbox read-only

⏺ Before relying on Codex's points, I'll check the main ones against the code.
Ran 1 shell command
⏺ They hold up. I'll apply the fixes now.
⏺ Update(README.md)
⎿ Updated README.md (+46 −17)

How it works

Two models, one terminal.

  1. You work in Claude Code as usual.Nothing changes about how you use it.
  2. Claude calls Codex when it helps.Subtle bugs, risky diffs, disputed design calls — or whenever you say “ask GPT”. You can also run /connect:ask yourself.
  3. Codex reads your code, read-only.It looks at the real files, so nothing needs to be pasted, and it can't change your tree.
  4. Claude checks the answer before acting.Codex's reply is treated as a second opinion, not as fact. The thread is saved, so you can continue it later.

Install

Two commands.

You need Claude Code, Node.js 18+, and the standalone Codex install signed in to your ChatGPT account. macOS or Linux.

terminal
claude plugin marketplace add Gary0302/connect
claude plugin install connect@connect-local

Restart Claude Code, then run /connect:doctor to check that everything is connected.

/connect:ask <question>send a question to Codex yourself
/connect:threadslist earlier Codex conversations
/connect:resume <id> <question>continue one of them
/connect:agy <question>ask Google's Antigravity CLI instead, as a third view
/connect:configpick the Codex model and reasoning effort
/connect:doctorcheck the setup
Full documentation on GitHub ↗

Not on Claude Code?

Use the standalone skill.

ask-codex is a plain SKILL.md plus one script around codex exec. No plugin, no daemon. It works in any agent that loads skills and can run a shell command, such as Cursor, Gemini CLI or Codex itself. It keeps intent-based routing and resumable threads.

terminal
git clone --depth 1 https://github.com/Gary0302/connect /tmp/connect
cp -r /tmp/connect/portable/ask-codex ~/.claude/skills/

Swap ~/.claude/skills/ for your agent's skills folder. You need Node.js 18+ and the codex CLI signed in. Unlike the plugin, it can't stop a turn mid-way once it exceeds its token ceiling; it can only flag it afterwards.

Skill source on GitHub ↗