Blog
Every AI model, inside the app you already use.
Setup guides for each app, how the model leaderboard works, and notes from past experiments.
Setup guides
One link, one sign-in. Ask GPT from Claude, Claude from ChatGPT, or Gemini from Codex.
Guide · November 2026
Add Polydev to Claude and ChatGPT
Add one link as a custom connector in Claude (web, desktop and mobile) or a custom MCP server in ChatGPT, sign in once, and ask GPT from Claude or Claude from ChatGPT.
Guide · November 2026
Using Polydev with Claude Code
One command adds the hosted link; sign in with /mcp. Then ask GPT, Gemini or Grok from Claude Code. The local client adds images and your ChatGPT plan.
Guide · November 2026
Using Polydev with Codex
Add the hosted link with codex mcp add, sign in with codex mcp login, and ask Claude or Gemini from Codex. The local client adds Nano Banana and GPT Image.
Guide · November 2026
Using Polydev with Cursor
Add Polydev with one click or one line in mcp.json, connect, and ask any model from Cursor chat.
Guide · November 2026
Using Polydev with Cline
Add Polydev as a remote MCP server in Cline, sign in in the browser, and ask any model from VS Code.
Guide · November 2026
Using Polydev with Windsurf
Add one entry to Windsurf's MCP config, sign in, and ask any model from Cascade.
Leaderboard and research
Deep dive · February 2026
How the AI Model Leaderboard Works
Elo ratings, pairwise comparisons, rankings from rank_perspectives, and prompt labels across 6 dimensions. The leaderboard keeps only labels and a one-way hash of each prompt.
Guide · February 2026
How Polydev Users Feed the Model Leaderboard
What each ask_model call records, how to rank the answers with rank_perspectives, and how to filter the public rankings for your stack.
Open source · Updated November 2026
The Polydev Local Client Is Open Source Under Apache-2.0
The local client, polydev-ai on npm, is published under Apache-2.0. What that covers and how to run it.
Research · January 2026
Multi-Model Consultation on SWE-bench Verified: A January 2026 Experiment
A past experiment: Claude Haiku 4.5 on 500 real GitHub issues, alone and with two consultant models. Results reflect that setup, not the current product.
Switch models, not apps.
Add one link to the app you already use and sign in once. 10 free requests to try. After that, your ChatGPT plan or your own API keys pay.