v1.1.2

Highlights

• Fixed project-scoped Codex tasks against the new multi-environment envelope by selecting the trusted primary workspace instead of rejecting valid project turns.
• Fixed permission-gated ChatGPT app/MCP calls: manual Allow or Deny now pauses and resumes the same browser turn, and unanswered prompts are denied after 60 seconds instead of terminating the run.
• Fixed mid-run steering so queued user instructions start a context-complete provider round while retries and tool-result continuations still reuse the active session.
• Fixed the first-time ChatGPT capability inspection timeout mismatch and now surfaces the real verification failure instead of a generic operation-aborted error.
• Suppressed only proven superseded Electron navigation aborts; genuine browser initialization failures still fail explicitly.
• Fixed setup for Codex configurations that explicitly declare the built-in model_provider = "openai"; genuine foreign routing conflicts still fail closed and remain reversibly replaceable.
• Added account-aware ChatGPT Web availability: Plus exposes Instant, Medium, and High, while Extra High and Pro remain limited to Pro-capable accounts.
• Added account-specific context windows and automatic compaction thresholds: 225K / 202.5K for Plus and 256K / 230.4K for Pro.
• Added a native context_length_exceeded preflight with clear guidance to switch models or run /compact before retrying.
• Detects ChatGPT's terminal "Something went wrong" alert immediately and returns a retryable upstream error instead of leaving the turn stalled.
• Added English and Chinese setup guidance reminding users to create a ChatGPT workspace for tunnel setup.
• Expanded regression coverage across model routing, context limits, browser failures, compaction, setup, permissions, project environments, and steering.

Verification

• Full core and launcher test suites
• TypeScript type checks and dependency audit
• Production renderer and relocatable runtime builds
• Packaged launcher smoke tests on macOS Intel, macOS Apple Silicon, Windows, and Linux
• macOS embedded-runtime and application signature verification
• Every published asset matched the release checksum manifest

Full Changelog: v1.1.1...v1.1.2

v1.1.1

Highlights

• Fixed ChatGPT rate-limit dialogs hijacking keyboard-based effort selection. The integration now recognizes the specific "Too many requests / You're making requests too quickly" dialog, acknowledges it, and returns a real retryable 429 rate_limit_error immediately instead of waiting about 70 seconds.
• Prevented effort-menu and other typed adapter failures from being mislabeled as model capacity errors. Structured error metadata now survives the launcher browser-helper boundary, including on Windows.
• Removed the hard 40-minute timeout for ChatGPT Pro turns. Active Pro turns now continue until completion or explicit cancellation unless a timeout is configured.
• Fixed codex_bind_turn for Web-model turns without a configured deadline. Full-harness tool calls now remain bound until the turn completes, is cancelled, or its broker connection closes, and binding returns expires_at: null instead of failing with Invalid Date (#29, thanks @Albro3459).
• Restored network access for routed Codex sessions by forwarding standalone Web Search through the authenticated native /v1/alpha/search endpoint.
• Fixed sandbox detection for Codex 0.146+ permission profiles, including read-only, workspace-write, and danger-full-access modes.
• Fixed subagent routing for every ChatGPT Web model. All five Web models are registered as valid local Responses models, while the integration keeps the readable MultiAgent V1 transport and restores previous user settings on uninstall.
• Upgraded the bundled runtime from Bun 1.3.11 to Bun 1.3.14 on every platform, with an exact build-time version guard and synchronized runtime notices.
• Fixed the Windows PowerShell installer so a successful NSIS installation is verified and launched from the actual codex-web-gpt-launcher directory.
• Improved stability across network transport, sandbox handling, mutable Markdown rendering, long-running turn brokers, compaction with images, Electron browser tabs, Windows named pipes, IPC, and child-process teardown.

Verification

• Full core and launcher test suites
• TypeScript type checks and dependency audit
• Production renderer and relocatable runtime builds
• Packaged launcher smoke tests on macOS, Windows, and Linux
• Windows PowerShell 5.1 installer contract
• Embedded Bun 1.3.14 version and runtime manifest checks
• Regression coverage for the exact rate-limit dialog, unrelated dialogs, launcher-helper error transport, and response classification

The v1.1.1 assets were rebuilt from the refreshed v1.1.1 tag and replace the earlier bundles.

Full Changelog: v1.1.0...v1.1.1

1.1.0

Codex Web GPT 1.1.0

ChatGPT Web can now run up to five Codex tasks at the same time.

Parallel browser tabs

• Each Codex task owns an independent launcher tab and an exact browser surface.
• Parallel tasks and subagents open new tabs instead of replacing the active ChatGPT turn.
• Later provider and tool rounds return to the same task tab.
• Closing a running tab terminates that exact browser turn.
• The five-tab limit is intentional to avoid excessive parallel traffic on the ChatGPT account.

Pro subagents and complete responses

• All ChatGPT Web models use Codex MultiAgent V1, so Web-rooted tasks can pass a plaintext task and forked context to ChatGPT Web Pro.
• Opaque V2 cross-backend payloads now fail explicitly instead of opening an empty Pro task.
• Multi-root ChatGPT answers are aggregated completely instead of returning only the final fragment.

Tested

• 167 core tests and 134 launcher tests.
• Release packages were built and smoke-launched on Windows, macOS Intel, macOS Apple Silicon, and Linux.

1.0.2

Codex Web GPT 1.0.2

ChatGPT Pro can now be spawned as a native Codex subagent.

Pro as a native consultant

A normal tool-capable agent can collect evidence from the computer and then ask Pro for a long, deep second opinion. With fork_turns="all", the Pro child receives the parent task's retained context snapshot, works inside the original Codex Harness, and returns its conclusion to the parent agent.

This is especially useful because ChatGPT Pro cannot initiate local computer tools itself. Pro remains read-only for the local machine; the change only removes the hardcoded catalog restriction that previously kept it out of Codex's native subagent registry.

Unlike projects that expose ChatGPT as an MCP query tool, this does not ask a separate browser tab for an isolated answer. Pro is the actual model of a native Codex child agent, so Codex owns the fork, context handoff, lifecycle, trace, and result collection. No second ChatGPT-query bridge is needed beyond the normal Codex Web GPT setup.

Stability and recovery

• Improved launcher and runtime stability on Windows and macOS.
• Added a persistent bridge switch in Launcher Settings.
• Disconnecting restores the previous native Codex backend and disables the ChatGPT Web route.
• Reconnecting starts the bridge again and makes the Web models available. A disconnected bridge stays disconnected after a launcher restart until you reconnect it.

The native-provider passthrough follows the same basic routing principle as OpenCodex: native Codex models still reach the native backend unchanged, and the original model catalog is preserved. The recovery switch therefore does not require sacrificing normal Codex behavior.

Tested

The launcher is manually tested on Windows 11 and macOS. The release pipeline also builds, verifies, packages, and smoke-launches Linux, although I still do not own a Linux machine, so real Linux reports are especially welcome.

What comes next

Stability remains the first priority. Please open an issue if anything behaves incorrectly — I am taking this project seriously and will keep maintaining it.

Next I will keep improving the Pro workflow and explore native support for more providers. Contributions, issues, and forks are welcome. If this project helps you, I would really appreciate a star.

v1.0.1

Codex Web GPT 1.0.0

Hey everyone — thank you so much for the stars.

For the last three days I have been building the launcher so Codex Web GPT is genuinely easy to install and use, not merely technically possible.

Before this release, a complete setup took around 15 minutes and several manual terminal steps. With the launcher, it takes about two minutes and a few clicks. Installing the launcher itself takes one command.

What changed

• Native desktop launcher for macOS, Windows, and Linux
• Embedded ChatGPT browser, so login and every Web model turn stay in one visible place
• Guided Codex and optional MCP setup with short video instructions
• Structured runtime logs and clear errors
• Automatic local runtime and tunnel supervision
• The original Codex model catalog and harness remain intact
• ChatGPT Web tiers, including Pro when your account provides it, appear natively in Codex

The old version launched a separate Playwright browser that was harder to observe and easier to break. The new launcher owns the browser surface, setup, logs, and runtime lifecycle in one app. ChatGPT Web now feels like a native Codex model instead of a bridge glued onto the side.

Tool-capable ChatGPT tiers can use the complete Codex harness through the optional MCP connector. ChatGPT Pro retains its ChatGPT-native capabilities, but it cannot currently initiate local MCP calls; you can collect local context with a tool-capable tier and then switch to Pro.

This is not OpenCodex

OpenCodex is a broad router for many third-party model providers. Codex Web GPT is deliberately focused: it makes the ChatGPT Web tiers already included in your ChatGPT subscription available inside the original Codex harness, with local Codex task history, context, compaction, streaming, images, approvals, and optional MCP tools.

Tested

The launcher was manually tested on Windows 11 and macOS, including real 48,000-character tool-capable turns. The release pipeline also builds and smoke-tests Linux, but I do not own a Linux machine yet, so real Linux reports are especially welcome.

What comes next

Stability comes first. Please open an issue if anything behaves incorrectly — I am taking this project seriously and will keep maintaining it.

After stabilization, I plan to explore support for other providers such as Gemini. I will also keep working on the ChatGPT Pro tool story: either native MCP support if the product allows it, or a clean way to compensate for that limitation without degrading the Codex experience.

I believe useful AI should be available to as many people as possible. If you already pay for ChatGPT and use the services included in your subscription within the applicable rules, you should be able to use that access in the workflow that helps you most. This project makes that workflow as convenient and native as possible.

Contributions, issues, and forks are welcome. If this project helps you, I would really appreciate a star.

Codex Web GPT 1.0.0 — Native desktop launcher

Codex Web GPT 1.0.0

Hey everyone — thank you so much for the stars.

For the last three days I have been building the launcher so Codex Web GPT is genuinely easy to install and use, not merely technically possible.

Before this release, a complete setup took around 15 minutes and several manual terminal steps. With the launcher, it takes about two minutes and a few clicks. Installing the launcher itself takes one command.

What changed

• Native desktop launcher for macOS, Windows, and Linux
• Embedded ChatGPT browser, so login and every Web model turn stay in one visible place
• Guided Codex and optional MCP setup with short video instructions
• Structured runtime logs and clear errors
• Automatic local runtime and tunnel supervision
• The original Codex model catalog and harness remain intact
• ChatGPT Web tiers, including Pro when your account provides it, appear natively in Codex

The old version launched a separate Playwright browser that was harder to observe and easier to break. The new launcher owns the browser surface, setup, logs, and runtime lifecycle in one app. ChatGPT Web now feels like a native Codex model instead of a bridge glued onto the side.

Tool-capable ChatGPT tiers can use the complete Codex harness through the optional MCP connector. ChatGPT Pro retains its ChatGPT-native capabilities, but it cannot currently initiate local MCP calls; you can collect local context with a tool-capable tier and then switch to Pro.

This is not OpenCodex

OpenCodex is a broad router for many third-party model providers. Codex Web GPT is deliberately focused: it makes the ChatGPT Web tiers already included in your ChatGPT subscription available inside the original Codex harness, with local Codex task history, context, compaction, streaming, images, approvals, and optional MCP tools.

Tested

The launcher was manually tested on Windows 11 and macOS, including real 48,000-character tool-capable turns. The release pipeline also builds and smoke-tests Linux, but I do not own a Linux machine yet, so real Linux reports are especially welcome.

What comes next

Stability comes first. Please open an issue if anything behaves incorrectly — I am taking this project seriously and will keep maintaining it.

After stabilization, I plan to explore support for other providers such as Gemini. I will also keep working on the ChatGPT Pro tool story: either native MCP support if the product allows it, or a clean way to compensate for that limitation without degrading the Codex experience.

I believe useful AI should be available to as many people as possible. If you already pay for ChatGPT and use the services included in your subscription within the applicable rules, you should be able to use that access in the workflow that helps you most. This project makes that workflow as convenient and native as possible.

Contributions, issues, and forks are welcome. If this project helps you, I would really appreciate a star.

Pinned Windows Bun runtime bbe3f6a26

Pinned Windows runtime used to avoid the Bun 1.3.x localhost proxy crash. Source: oven-sh/bun canary revision bbe3f6a26. Archive SHA-256: fba998226ac3f8ce75fc5c108a622deb96ded18e72918c26693ab8aca36adebb.

v0.1.16

Full Changelog: v0.1.15...v0.1.16

v0.1.15

Full Changelog: v0.1.14...v0.1.15

v0.1.14

What's Changed

• Build native ChatGPT Web harness by @miuuyy in #8

New Contributors

• @miuuyy made their first contribution in #8

Full Changelog: https://github.com/miuuyy/codex-chatgpt-web/commits/v0.1.14