claudeers.
// Integrations & Connectors

claude-for-chatgpt-desktop

Use Claude Fable, Opus, Sonnet and Haiku in the ChatGPT desktop app on macOS, in the model picker next to GPT, with the app's own tools.

// Integrations & Connectors[ cli ][ api ][ desktop ][ web ][ claude ]#claude#chatgpt#claude-code#codex#macos#integrations◷ MIT$open-sourceupdated about 1 month ago
Actively maintained
99/100
last commit 10 days ago
last release none
releases 0
open issues 0
// star history

Install with your AI

Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.

Install and set up claude-for-chatgpt-desktop (git-clone project) into my current project.
Found on https://claudeers.com/claude-for-chatgpt-desktop
Repo: https://github.com/wuchris-ch/claude-for-chatgpt-desktop
Homepage/docs: —
Detected install method: git-clone → git clone https://github.com/wuchris-ch/claude-for-chatgpt-desktop
Category: integrations. Platforms: cli, api, desktop, web.
Read the repo's README for exact setup and env vars, then install it and wire it into my project.

Claudeers Health Verdict:
active; community-verified: false. Confirm the source before running anything.
// or clone
git clone https://github.com/wuchris-ch/claude-for-chatgpt-desktop

// compatibility

Platformscli, api, desktop, web
Operating systems—
AI compatibilityclaude
LicenseMIT
Pricingopen-source
LanguageJavaScript

Get your FREE $2.50 API credits to access TickAtlas financial data ↗

Claude for ChatGPT Desktop

Claude in the ChatGPT desktop app's model picker, right next to GPT. For macOS.

Illustrated walkthrough, redrawn from the app: choosing Claude Fable 5.1 in the model picker, Claude running the tests and reading a page in the in-app browser, and a message sent while it works

  • Pick Claude Fable, Opus, Sonnet or Haiku in any chat.
  • Claude uses the app's own tools: the in-app browser, computer control, terminal, file edits and plugins.
  • Switch between Claude and GPT in the same chat.
  • Runs on your own Claude Code login. GPT keeps its own login and features; its requests pass through the bridge on your Mac.

Full 50-second illustrated demo: docs/demo.mp4

Unofficial. Not affiliated with Anthropic or OpenAI. Usage counts against your Claude plan, like any Claude Code session (Anthropic's notes).

Install

You need

Run

curl -fsSL \
  https://raw.githubusercontent.com/wuchris-ch/claude-for-chatgpt-desktop/main/install.sh \
  | bash

Then pick a Claude model in ChatGPT. The installer offers to reopen ChatGPT for you.

  • Update: run the same command again. An update keeps the picker or the separate window, whichever you use; --picker or --window switches.
  • Remove: add -s -- --uninstall after bash.
  • Leave your normal ChatGPT untouched: add -s -- --window after bash for a separate Claude-only window.
What the installer does
  1. Checks for Node.js 24 or later (and offers to install or upgrade it with Homebrew) and for your Claude Code sign-in.
  2. Downloads this repository to ~/Library/Application Support/Claude for ChatGPT Desktop/source.
  3. Installs a small background service that listens only on 127.0.0.1.
  4. Backs up your ChatGPT settings and adds one setting, openai_base_url.
  5. Offers to reopen ChatGPT.

To read it before running it, clone the repository and run ./install.sh.

Install step by step instead
git clone https://github.com/wuchris-ch/claude-for-chatgpt-desktop.git
cd claude-for-chatgpt-desktop
npm ci
python3 scripts/setup.py
python3 scripts/install.py
python3 scripts/picker.py install

install.py installs the background service. picker.py install backs up your ChatGPT settings and adds the one setting. Then reopen ChatGPT. To undo it, run python3 scripts/picker.py uninstall and reopen ChatGPT.

Separate window with its own AGENTS.md and skills

The separate window (--window) has its own ChatGPT profile, so it can have its own instructions and skills:

  • Put an AGENTS.md file and a skills folder in ~/Library/Application Support/Claude for ChatGPT Desktop/codex/. Setup never replaces them.
  • Or share your usual ones: python3 scripts/setup.py --agents-md codex --share-skills links ~/.codex/AGENTS.md and ~/.codex/skills.

To open the window later, double-click Open ChatGPT with Claude.command in the downloaded folder, or run python3 scripts/launch.py. Sign into ChatGPT once in that window.

Features

Works like GPT does in the app

  • The in-app browser, computer control, terminal, file edits, plugins and connectors, with the app's approvals and tool cards.
  • Images, edits, regenerate, helper agents, review mode, chat titles and inline visuals.
  • Chats you drag in or @-mention, messages between threads, goals and automations.
  • Web search with your ChatGPT sign-in, shown in the chat like GPT's searches, with linked sources.
  • Message Claude while it works, even mid-reply. It answers you before its next step and carries on.
  • Stop, then continue where it left off.
  • A question card the app can't show goes back to Claude with the reason, so it asks in a form you can see.

Switch models freely

  • Claude and GPT in the same chat. Claude sees what GPT did, and GPT can read Claude's summaries.
  • Each Claude entry carries the version Claude Code runs, such as Claude Opus 5.5.
  • GPT keeps its own login, connection and compaction.

Fast and cheap on long chats

  • One Claude Code session per chat, resumed each turn and kept running through its tool calls. In daily use, 97 to 98% of input tokens came from the prompt cache.
  • Compaction reuses the cache, even when it fires in the middle of a turn.
  • Chats full of screenshots stay under the request size limit, and a tool call the app never answers can't hold a session forever.

Tested

  • 223 automated tests, plus live checks with real Claude Code, real OpenAI and the app's own runtime. The measurements behind each choice are in VERIFICATION.md.

Limits

  • macOS only. Tested with ChatGPT 26.1007.71326 (Codex runtime 0.162.0-alpha.17.2), Claude Code 2.1.300 and Node.js 24.21. A later app or Claude Code release can need an update here.
  • Local tasks only. Cloud tasks, image generation and other OpenAI-hosted features stay with GPT. Voice conversations run on OpenAI's voice model, which hands tasks to the selected Claude model. Other tools that run on OpenAI's servers are left out of Claude's tools.
  • GPT goes through the bridge in picker mode. If the service stops, GPT requests fail until macOS restarts it or you uninstall. The setting is in the shared ~/.codex/config.toml, so the Codex CLI and IDE extension also go through the bridge. The uninstall command works even when the service or its folder is gone. Picker mode needs a ChatGPT sign-in, doesn't use proxy settings, and isn't meant for ChatGPT workspaces with data-residency routing.
  • GPT's summaries are encrypted for GPT. If a chat that GPT compacted switches to Claude, Claude gets the messages after the summary and a note that it's unavailable.
  • Helper agents whose model family differs from their parent's are untested.
  • Some request options are refused with an error instead of being ignored: forced tool choice, output-token caps and non-JSON text formats.
  • Long chats use your plan quickly. Each Claude entry has a 1M-token context and compacts at 800k tokens, and every tool step sends the whole context, mostly from the prompt cache. The cache figures above come from subscription use, where Claude Code caches prompts for an hour. To compact sooner, set a smaller context_window in a models file (see Models under Configuration).
  • Claude Code's own setup is not used. The bridge gives Claude the app's tools and prompt instead of Claude Code's, ignores Claude Code's settings files, and clears ANTHROPIC_* variables, so Bedrock, Vertex, apiKeyHelper and custom endpoint setups don't apply.
  • No silent model switches. If Claude Code's safety classifiers re-run a request on a fallback model, such as Opus 4.8 for Opus 5.5, the bridge stops the turn and says why.
  • Long browser runs drop older screenshots from Claude's context in batches (the app keeps them), so Claude takes a fresh screenshot when it needs one.

How it works

flowchart LR
    subgraph app["ChatGPT desktop app"]
        direction TB
        Picker["<b>Model picker</b><br/>GPT and Claude models"]
        Tools["<b>The app's tools</b><br/>computer use, in-app browser,<br/>terminal, plugins"]
    end
    Bridge["<b>Bridge</b><br/>127.0.0.1 on your Mac"]
    OpenAI["<b>OpenAI</b><br/>with the app's own login"]
    Claude["<b>Claude Code</b><br/>one resumed session per chat"]
    Picker -->|"model request"| Bridge
    Bridge -->|"GPT models"| OpenAI
    Bridge -->|"Claude models"| Claude
    Claude -.->|"tool call over MCP"| Bridge
    Bridge -.->|"app runs it, with approval"| Tools
  • One setting. Picker mode sets openai_base_url, so the app's OpenAI connection goes to the bridge, which adds the Claude models to the app's model list.
  • GPT requests go on to OpenAI as the app sends them, with the app's own login, WebSocket connection and compaction. In a chat where Claude also answered, the bridge turns Claude's summaries into text and drops Claude's item ids, which OpenAI can't read.
  • Claude requests are answered by the bridge: one headless Claude Code session per chat, resumed each turn, with the app's tools offered over MCP.
  • Tool calls go back to the app as ordinary tool calls. The app applies its approvals, runs the tool and returns the result to Claude.
  • Nothing is patched. The app bundle isn't modified, and the bridge has no tools of its own. The separate window works the same way through its own model provider, without the OpenAI relay.

Compared with similar projects

Checked October 14, 2026. Stars are GitHub stars on that date.

ProjectHow Claude runsClaude uses the app's browser and computer controlNotes
This projectBridge on the app's model API; claude -p per turn, kept running through its tool calls and resumed between turns; the app's tools over MCPYesBuilt only for the ChatGPT desktop app. Claude next to GPT in the normal picker with GPT traffic relayed to OpenAI, or a separate window
yetone/magpie (12.7k stars)Model gateway for many agents and providers; Claude plan requests run the local claude with the client's tools over MCPNot documented for the desktop appAdds its models next to GPT in the Codex picker through openai_base_url. A menu bar app that also routes Claude Code, Gemini CLI, OpenCode and other agents to any provider
gkorepanov/ccodex (33 stars)Claude Agent SDK behind its own codex app-server gatewayNo; Claude uses Claude Code's own toolsReleased almost daily. Claude next to GPT in the picker, model switching mid-chat, plan mode, /goal, side chats, sub-agents, the phone app, SSH servers and Linux
EthanSK/claude-in-codex (0 stars)claude -p per message with --resume; the app's tools over MCPComputer use tested through Codex CLI; its README says the in-app browser wasn't testedInstalled from a clone
Reidond/codex-claude-models-plugin (3 stars)Claude Agent SDK as a model providerNot documentedNo images or compaction for Claude; no updates since September 14
wbopan/claude-in-codex (0 stars)Menu bar app that hooks the app's internals through the Node inspectorNot documentedNotarized DMG; app version 26.924 disabled the hook
lidge-jun/opencodex (17.4k stars), router-for-me/CLIProxyAPI (55.2k stars)General model proxiesOnly through raw OAuth tokens or an API keyWith raw Claude OAuth tokens: Anthropic blocked third-party use on January 9, 2026 and has billed it as extra usage since April 4, 2026. Their claude CLI mode is stateless and has no tools

FAQ

Can this get my Claude account banned?

The bridge runs the official, unmodified Claude Code CLI (claude -p) with your own login. It never reads, copies or forwards your Claude credentials, and it doesn't offer a Claude login of its own. Anthropic says you can still use claude -p and third-party apps with your plan's usage limits (update of October 7, 2026). For heavy use, switch to an Anthropic API key (see "Claude login or API key" under Configuration).

Does it change anything for GPT?

In picker mode, GPT requests pass through the bridge on your Mac and go on to OpenAI as the app sent them, with the app's own login, WebSocket connection and compaction. Picker mode changes one setting, openai_base_url, and --uninstall takes it back out. In a chat where Claude also answered, the bridge drops Claude's item ids before relaying, because OpenAI rejects ids it didn't create.

Why not just use Claude Code or the Claude app?

This gives Claude the ChatGPT app's tools and interface: the in-app browser, computer control, plugins and connectors, approvals and tool cards. You can also switch between Claude and GPT in the same chat.

Why not just ask Claude or Codex to build one?

You can, and a basic bridge works in an afternoon. The hard part is the long tail that only shows up in daily use: keeping Claude's prompt cache warm (97 to 98% of input from the cache), compaction that doesn't resend the whole chat, Stop and resume without losing tool results, screenshots that push a long chat past the size limit, and GPT chats that keep working after Claude has answered in them. Its separate-window form has been used daily for over two weeks, and VERIFICATION.md lists each problem found, its fix and the test that covers it.

Configuration

Run the commands below from the downloaded folder (~/Library/Application Support/Claude for ChatGPT Desktop/source) or your clone. python3 scripts/setup.py --help lists every option. Setup rewrites the Claude profile's settings each time it runs, then python3 scripts/install.py applies them to the service.

Picker mode
  • python3 scripts/picker.py install adds a marked openai_base_url line to config.toml in your ChatGPT profile (~/.codex, or $CODEX_HOME), pointing at the bridge with a private key in the path.
    • It first saves config.toml and the app's cached model list to backups/ in the runtime folder.
    • It refuses to run if you already set openai_base_url yourself or the profile uses another model provider.
  • python3 scripts/picker.py uninstall restores config.toml byte for byte when nothing else in it changed. Otherwise it removes only the marked line and keeps your later edits. It also restores the cached model list. If the runtime folder is gone, it still removes the marked line.
  • A config.toml that is a symlink, such as one kept in a dotfiles repository, stays a link; the file it points to is changed.
  • python3 scripts/picker.py status shows whether picker mode is on.

Reopen ChatGPT after either change.

Models

claude-models.json lists the picker entries.

  • Each claude_model is passed to claude --model.
  • The aliases fable, opus, sonnet and haiku follow Claude Code to new releases. On October 9, 2026 they started claude-fable-5-1, claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5.
  • An alias entry is named after the version Claude Code last started for it, so Claude Fable 5.1 becomes Claude Fable 5.5 after the first chat on the new release.
  • The bridge checks the model Claude Code actually starts. An alias accepts any version of its family; a full model id must match exactly.

To pin a version or add a model, copy the file, edit it, and pass it to setup:

{"slug": "claude-opus-5-5", "claude_model": "claude-opus-5-5", "display_name": "Claude Opus 5.5",
 "efforts": ["low", "medium", "high", "xhigh", "max"], "default_effort": "medium", "context_window": 1000000}
python3 scripts/setup.py --models ~/my-claude-models.json
Claude login or API key
  • By default the bridge runs the unmodified claude binary with your normal Claude Code login.
  • It never reads Claude Code's credentials, and it clears inherited ANTHROPIC_* and CLAUDE_* variables (except CLAUDE_CONFIG_DIR) so nothing else changes the account or endpoint. Bedrock, Vertex and custom endpoint setups therefore don't apply.
  • For heavy use, an Anthropic API key is the better fit: Anthropic's plan limits assume ordinary, individual use of Claude Code. In API key mode Claude Code uses the key instead of your login.
python3 scripts/setup.py --auth api-key
ANTHROPIC_API_KEY=your-key python3 scripts/install.py

The key is stored only in the service definition, ~/Library/LaunchAgents/<label>.plist, written with mode 0600, and kept when you reinstall. Switch back with --auth claude-login.

Web search

In the picker, Claude searches the web whenever the app's web search is on, which follows your ChatGPT settings (web_search in ~/.codex/config.toml: cached by default, or live).

  • Claude gets a web_run tool with the same commands as the app's own: searching, opening and finding text in pages, plus finance, weather, sports and time.
  • The bridge sends each search to OpenAI's search endpoint (https://chatgpt.com/backend-api/codex/alpha/search) with the sign-in the app sent with that request: only the Authorization, ChatGPT-Account-Id and originator headers, to that one URL, with redirects refused. Like the app, it adds your last two messages and up to about 1,000 tokens of the reply between them, for context.
  • The sign-in stays in memory. The bridge logs only status codes and command types.

The separate window needs it turned on:

python3 scripts/setup.py --web-search
python3 scripts/install.py
  • This turns on the app's built-in web.run tool for Claude.
  • The bridge forwards each search to OpenAI's search endpoint (https://chatgpt.com/backend-api/codex/alpha/search) with the Claude window's own ChatGPT sign-in: only the Authorization, ChatGPT-Account-Id and originator headers, to that one URL, with redirects refused.
  • Searches can include conversation context that the app adds. The bridge reads no credential files and logs only status codes and command types.
Other options
OptionDefaultEffect
--port19480Local port of the bridge. With picker mode on, turn it off first (picker.py uninstall), then on again after the change
--labelio.github.wuchris-ch.claude-for-chatgpt-desktopLaunchAgent label. Same rule as --port with picker mode on
--agents-mdnonecodex links ~/.codex/AGENTS.md into the Claude profile; claude copies ~/.claude/CLAUDE.md once
--share-skillsoffLinks ~/.codex/skills into the Claude profile
--app/Applications/ChatGPT.appThe ChatGPT app to open

Set CLAUDE_BRIDGE_RUNTIME to keep everything in another folder.

Files

Everything lives in ~/Library/Application Support/Claude for ChatGPT Desktop/:

PathContents
source/The copy of this repository that install.sh downloads and updates
app/Installed copy of the bridge
codex/The Claude window's profile: settings, model catalog, plugins and task history
user-data/The Claude window's browser and app state
sessions/Native session records, system prompts and an encrypted retry cache that expires after 24 hours
logs/Request, model, tool-name and lifecycle metadata, and service errors. An install starts a new log once it passes 20 MB
launch.json, token, gateway-keySettings from setup, and the local keys between the app and the bridge
picker.json, backups/Picker mode's record of what it changed, and the backups it restores

Claude Code keeps its own transcripts of these chats in ~/.claude/projects/, in the folder whose name ends in Claude-for-ChatGPT-Desktop-sessions-empty-workspace. Uninstalling, even with --delete-data, leaves them.

Security and privacy

  • Local only. The bridge listens only on 127.0.0.1, requires a random local key, and refuses requests from browsers. The keys are kept in owner-only files, never on a command line. Picker mode makes config.toml owner-only while its setting holds the key, and restores the file's mode when turned off.
  • Separate logins. Claude Code manages its own login. The app's ChatGPT credentials never reach Claude. In picker mode the bridge passes them, with the rest of each GPT request, only to OpenAI at chatgpt.com.
  • No content in logs. Logs hold request, model, tool-name and lifecycle metadata, plus the last lines of Claude Code's error output when a turn fails. They never contain prompts, tool arguments, tool results, tokens or keys.
  • One setting changed. Apart from the one setting picker mode adds, your usual ChatGPT profile in ~/.codex is only read, to copy its settings and plugins into the separate window's profile.

Uninstall

curl -fsSL \
  https://raw.githubusercontent.com/wuchris-ch/claude-for-chatgpt-desktop/main/install.sh \
  | bash -s -- --uninstall

Or from a clone:

python3 scripts/picker.py uninstall          # take Claude out of the normal picker
python3 scripts/uninstall.py                 # also stop and remove the service, keep the Claude profile
python3 scripts/uninstall.py --delete-data   # also delete the Claude profile, history and logs (not Claude Code's transcripts, see Files)

A chat where Claude answered may not continue on GPT once the bridge is removed, because OpenAI can't read Claude's items in it. Start a new chat for those.

Design notes

Prompts, sessions, model switching and the WebSocket relay
  • Prompts. The app's base prompt is written for Codex. The bridge renames its identity lines for the selected Claude model and tells Claude to ignore Claude Code's own working-directory note. Developer messages before the conversation become the system prompt. Notes the app adds later, such as the date, the open page or an interrupted turn, reach Claude where they occur, so a note does not restart a tool cycle or invalidate the prompt cache.
  • Patches and host habits. GPT's apply_patch output is constrained by a grammar that MCP cannot carry, so the bridge adds a written patch format guide, plus notes on escaping inside JavaScript and on the shared output budget of an exec cell. It also passes on host habits found by benchmarking: render PDFs to images in a temporary folder, keep long jobs in the foreground, and avoid rm -rf, which the app refuses. When the app offers its automation tool, it adds how to write a one-time timer that the app's scheduler runs once, at the local time.
  • Sessions. Between user turns Claude Code resumes the saved session. Before resuming, the bridge checks the app's transcript against what the session saw. Edits, undo, regeneration and unexpected failures rebuild from the app's transcript. A turn stopped by the user, cut off by a bridge restart, ended by an API error such as an overload or a usage limit, or cut off when Claude Code exits while the app runs a tool resumes its native session; a Stop before Claude Code has started waits for it to start, so the message is recorded once. When the user edits or withdraws a stopped message whose turn sent no tool call, Claude Code resumes at the end of the last completed turn and drops the stopped one. While Claude thinks or writes a long tool call, the bridge sends a no-op event every 10 seconds so the app's stall timer does not fire.
  • Model switching. Switching Claude models between turns resumes the same native session on the new model; that first request cannot reuse the previous model's cache. If GPT answered in between, its messages and tool activity are passed to the resumed session as history. A compaction keeps the model the session ran on, so its cache still applies.
  • WebSocket relay. The bridge connects to OpenAI during the app's handshake, so the headers the app reads from it (turn routing, server model) arrive unchanged, then routes each response.create by model. For Claude it rebuilds the full input from the increments the app sends after the first request, answers prewarm requests without inference, and treats response.interrupt as a Stop. OpenAI closes connections that sit idle, as this one does while Claude answers; the app's side then stays open for Claude, and a GPT request on it makes the app reconnect. A Claude compaction summary is a sealed item only the bridge can open, so it is turned back into text before a GPT request leaves for OpenAI.

Development

npm test               # 201 tests with stand-ins for Claude Code and OpenAI
node e2e/live.mjs      # eighteen short checks with real Claude Code

License

MIT

// faq

What is claude-for-chatgpt-desktop?

Use Claude Fable, Opus, Sonnet and Haiku in the ChatGPT desktop app on macOS, in the model picker next to GPT, with the app's own tools.. It is open-source on GitHub.

Is claude-for-chatgpt-desktop free to use?

claude-for-chatgpt-desktop is open-source under the MIT license, so it is free to use.

What category does claude-for-chatgpt-desktop belong to?

claude-for-chatgpt-desktop is listed under integrations in the Claudeers registry of Claude-compatible tools.

19 views
★ 13 stars
unclaimed
updated about 1 month ago

// embed badge

claude-for-chatgpt-desktop on Claudeers
[![Claudeers](https://claudeers.com/api/badge/claude-for-chatgpt-desktop.svg)](https://claudeers.com/claude-for-chatgpt-desktop)

// retro hit counter

claude-for-chatgpt-desktop hit counter
[![Hits](https://claudeers.com/api/counter/claude-for-chatgpt-desktop.svg)](https://claudeers.com/claude-for-chatgpt-desktop)

// reviews

// guestbook

0/500

// related in Integrations & Connectors

🔓

Use claude code and codex for free in the terminal, VSCode extension, and discord like OpenClaw (voice supported)

// integrationsAlishahryar1/⟨Python⟩★ 72,248◷ MIT[ claude ]
🔓

Bridge local AI coding agents (Claude Code, Cursor, Gemini CLI, Codex) to messaging platforms (Feishu/Lark, DingTalk, Slack, Telegram, Discord, LINE, WeChat…

// integrationschenhg5/⟨Go⟩★ 16,316[ claude ]
🔓

All parts of Claude Code's system prompt, 27 builtin tool descriptions, sub agent prompts (Plan/Explore/Task), utility prompts (CLAUDE.md, compact, statusli…

// integrationsPiebald-AI/⟨JavaScript⟩★ 13,191◷ MIT[ claude ]
🔓

Claude Code skill/plugin: immersive scroll-scrubbed 'fly through the world' landing pages generated with Higgsfield (Emons-style diorama flights, seamless co…

// integrationsoso95/⟨JavaScript⟩★ 11,032◷ MIT[ claude ]
→ see how claude-for-chatgpt-desktop connects across the ecosystem