Polydev
Back to blog

Guide · Updated November 6, 2026

Using Polydev with Codex

Ask Claude, Gemini or Grok from inside Codex. Add the hosted link with two commands, sign in in the browser, and name the model you want. With the local client, Codex can also make images with Nano Banana or GPT Image.

What Polydev adds

Polydev is a tool Codex calls, not a replacement for its model. Codex keeps its own model and sign-in, and calls Polydev only when you name another model. The answer comes back in your session and says what paid for it. The hosted link has four tools:

ask_model

Sends a prompt to one model, or several, and returns each answer with what paid for it. The older name get_perspectives still works.

list_models

The models you can reach right now, and what pays for each.

search_x

Recent public posts on X. Read-only, and it needs your own xAI API key.

rank_perspectives

Ranks the answers, best first, when you ask. Rankings feed the public leaderboard as counts.

Set up the hosted link

Nothing to install and no keys in config files.

  1. 01

    Add the link

    terminal
    codex mcp add polydev --url https://polydev.ai/mcp
  2. 02

    Sign in

    Your browser opens the Polydev sign-in; approve access.

    terminal
    codex mcp login polydev
  3. 03

    Start Codex

    Start Codex as usual. Polydev's tools are now available in your sessions.

Optional: the local client

The hosted link is all most people need. The local client is the npm package polydev-ai (version 1.12.0, Node.js 20 or newer, open source under Apache-2.0). It runs on your machine and adds three things: image models, your ChatGPT plan, and the Claude Code and Codex you are already signed in to.

Called from Codex, it can run your own Claude Code under your own login. It never runs Codex itself, so Codex does not call itself.

Add this to ~/.codex/config.toml:

~/.codex/config.toml
[mcp_servers.polydev-local]
command = "npx"
args = ["-y", "polydev-ai@latest"]
tool_timeout_sec = 300

The first call opens your browser to sign in to Polydev (or ask Codex to run login). If you keep both the hosted link and the local client, give them different names, as above.

On top of the hosted tools, the local client has:

generate_image

Makes an image with an OpenAI or Google model, such as GPT Image or Nano Banana. Uses your key, or one free request.

connect_chatgpt_plan

Continue with ChatGPT: sign in once in the browser and GPT requests use your ChatGPT Plus or Pro plan.

chatgpt_plan_status

Shows whether your ChatGPT plan is connected.

login

Signs in to Polydev in the browser and saves the token on this machine.

get_auth_status

Shows the connection, the Claude Code and Codex it found, and your configured models.

To have GPT requests use your ChatGPT Plus or Pro plan, ask your app to run connect_chatgpt_plan, or sign in from a terminal:

terminal
npx -y polydev-ai chatgpt login     # Continue with ChatGPT in the browser
npx -y polydev-ai chatgpt status    # is plan usage on?
npx -y polydev-ai chatgpt models    # models your plan can use
npx -y polydev-ai chatgpt logout    # sign out and delete the tokens

Continue with ChatGPT works in the local client only for now. On the hosted link it is waiting for OpenAI's approval.

On a machine without a browser, the local client can read a Polydev token from the environment instead; see the docs.

What to ask

Ask in plain words and name the model, or mention Polydev. Codex calls Polydev when you ask, or may offer to; it does not call it on its own.

  • Check a plan

    “Before you change anything, ask Claude to list the risks in this migration plan.”

  • Compare two answers

    “Ask Claude and Gemini to review the patch you just wrote, then tell me where they disagree.”

  • Debugging

    “Ask Gemini why this function leaks memory under load. Include the profiler output.”

  • Make an image (local client)

    “Make a 1200×630 cover for this post with Nano Banana.”

Which models you can reach depends on what you set up in your dashboard. Ask for list_models to see them.

Who pays

Polydev is free: there are no plans, tiers or credits to buy. Each request uses the first of these you have set up, and every answer ends with what paid for it, for example Paid by: your API key.

Your ChatGPT plan

Local client. GPT requests count toward your Plus or Pro plan, with a weekly cap you set in ChatGPT.

Your Claude Code or Codex

Local client. Runs the Claude Code or Codex you are signed in to on this machine.

Your API keys

OpenAI, Anthropic, Google, xAI or OpenRouter, added in your dashboard. The provider bills you directly.

10 free requests

On Polydev, up to $1 in total. No card.

Troubleshooting

Polydev does not show up

Run codex mcp list. If Polydev is missing, run the add command again and restart Codex.

Calls fail with an authorization error

Run codex mcp login polydev again to sign in. You do not need to paste a token.

Slow answers

Several models take longer than one. A short, focused question with the relevant code comes back faster than a whole file.

A model is missing, or the free requests ran out

Add your own API keys or an OpenRouter key in your dashboard, or use the local client with your ChatGPT plan or your Claude Code. Ask for list_models to see what you can reach and what pays for each.

Want to disconnect

Remove the server from Codex, and revoke access under Dashboard → Connected apps.

Other apps

The same link works in Claude, ChatGPT, Claude Code, Codex, Cursor, VS Code, Windsurf and Cline. Prompts go only to the model that answers; they are never used for training or sold, and prompt and response text is deleted after 30 days. See the privacy policy.