The Codex IDE Extension: VS Code, JetBrains, and the Hybrid Cloud-Local Workflow

The Codex IDE Extension: VS Code, JetBrains, and the Hybrid Cloud-Local Workflow

The Extension Landscape

OpenAI ships Codex across four surfaces: the CLI, the standalone macOS app, Codex Cloud on the web, and the IDE extension. Of these, the IDE extension is the one most developers encounter first — and it is also the least documented in technical depth. This article covers the VS Code extension, the JetBrains integration, and the hybrid cloud-local workflow that makes the extension more than just a chat sidebar.

Installation and Compatibility

VS Code and Forks

The Codex extension is available from the Visual Studio Marketplace under the publisher openai 1. It works with VS Code itself and with forks including Cursor and Windsurf. Platform support covers macOS and Linux natively; Windows is experimental, and OpenAI recommends running Codex in a WSL workspace for the best Windows experience 1.

# Install via the VS Code CLI
code --install-extension openai.chatgpt

After installation, Codex appears in the right sidebar. If it does not, restart the editor. The extension prompts for authentication via either a ChatGPT account or an OpenAI API key. ChatGPT Plus, Pro, Business, Edu, and Enterprise plans all include Codex usage credits 2.

JetBrains Integration

Since January 2026, Codex is natively integrated into JetBrains IDEs — IntelliJ IDEA, PyCharm, WebStorm, Rider, and others — starting with version 2025.3 3. Three authentication methods are supported: a JetBrains AI subscription, a ChatGPT account, or an OpenAI API key 4.

The JetBrains integration embeds Codex within the existing JetBrains AI chat interface rather than shipping as a standalone sidebar. This means it shares the UI with other JetBrains AI providers, which trades some Codex-specific UX for consistency with the JetBrains ecosystem.

⚠️ The JetBrains integration was initially powered by GPT-5.2-Codex 4. Confirm your model selection is current — GPT-5.4 is now the recommended default 5.

The Three Agent Modes

The extension’s most important UX decision is the mode switcher beneath the chat input 6:

flowchart LR
    A[User Prompt] --> B{Mode?}
    B -->|Chat| C[Read-only discussion<br/>No file changes]
    B -->|Agent| D[Read + edit + run<br/>Within working directory]
    B -->|Agent Full Access| E[No approval required<br/>Full filesystem + network]

Chat Mode

Pure conversation. Codex can read files for context but makes no changes. Use this for planning, architecture discussion, or understanding unfamiliar code before committing to edits.

Agent Mode (Default)

Codex can read files, make edits, and execute commands within the working directory. Tool calls require approval by default — the extension shows a diff preview before applying changes. This maps to approval_policy = "on-failure" in config.toml 7.

Agent (Full Access)

Equivalent to --full-auto or approval_policy = "never" in the CLI. Codex runs without stopping for approval, with full filesystem and network access. The official documentation explicitly warns to exercise caution 6.

Configuration: Shared with the CLI

The IDE extension and CLI share the same ~/.codex/config.toml file 7. This is a critical architectural decision — any model, sandbox, or approval configuration you set for the CLI applies identically in the IDE, and vice versa.

# ~/.codex/config.toml — shared between CLI and IDE extension
model = "gpt-5.4"
approval_policy = "on-failure"
sandbox_mode = "workspace-write"

[profiles.deep]
model = "gpt-5.3-codex"
model_reasoning_effort = "high"

[profiles.fast]
model = "gpt-5.3-codex-spark"
model_reasoning_effort = "medium"

⚠️ One limitation: profiles (--profile) are not currently supported in the IDE extension 7. You must switch models manually via the model selector or change the top-level model key in config.toml. This is a known gap — the CLI’s profile system is still marked experimental.

Extension-Specific Settings

The IDE extension exposes a small set of VS Code-specific settings beyond config.toml 8:

Setting Purpose
chat.fontSize Controls text size in the Codex sidebar
chat.editor.fontSize Controls code/diff font size in conversations
chatgpt.commentCodeLensEnabled Enables CodeLens indicators above TODO comments
chatgpt.openOnStartup Focus the Codex sidebar on editor launch
chatgpt.runCodexInWindowsSubsystemForLinux Route execution through WSL on Windows
chatgpt.localeOverride Override UI language (auto-detects if empty)
chatgpt.cliExecutable Custom CLI binary path (development use only)

The chatgpt.cliExecutable setting exists for contributors actively developing the Codex CLI. Changing it in production may cause instability 8.

Commands and Keyboard Shortcuts

The extension registers six commands with VS Code 9:

Command Default Shortcut Action
chatgpt.newChat Cmd+N / Ctrl+N Create a new thread
chatgpt.addToThread — Add selected text as context
chatgpt.addFileToThread — Add entire file as context
chatgpt.implementTodo — Address a selected TODO comment
chatgpt.newCodexPanel — Open an additional Codex panel
chatgpt.openSidebar — Open the Codex sidebar

Customise shortcuts via Preferences: Open Keyboard Shortcuts (⌘⇧P → search for the command ID). The chatgpt.addToThread command is particularly useful when bound to a shortcut — select a code region, trigger the binding, and the selection becomes context for your next prompt without needing to type @filename.

File References

Tag files directly in prompts using @filename.tsx syntax 6. The extension resolves these against your workspace, providing the model with full file contents as context. This is more reliable than hoping the agent will find the right file — explicit references reduce hallucination risk.

Cloud Delegation: The Hybrid Workflow

The IDE extension’s most distinctive capability is seamless delegation between local and cloud execution 110.

sequenceDiagram
    participant Dev as Developer (IDE)
    participant Local as Local Agent
    participant Cloud as Codex Cloud

    Dev->>Local: Start conversation locally
    Dev->>Local: Interactive edits with approval
    Dev->>Cloud: "Run in the cloud" (heavy task)
    Cloud-->>Cloud: Sandboxed execution
    Cloud-->>Dev: Diff preview + PR
    Dev->>Local: Apply cloud changes locally
    Dev->>Local: Continue iteration

How It Works

  1. Start locally — chat, explore, make small edits with agent mode
  2. Delegate to cloud — click “Run in the cloud” for longer tasks. Choose whether the cloud environment runs from main (new ideas) or from your local changes (continuing work) 6
  3. Cloud executes in isolation — sandboxed containers, parallel execution, full network access
  4. Review and apply — preview the cloud diff in the IDE, ask follow-up questions, then apply locally

Context is preserved across the local-to-cloud handoff 6. When you continue locally after a cloud task, Codex remembers the full conversation history.

When to Delegate to Cloud

Cloud delegation is worth the context switch when:

  • The task will take more than a few minutes of agent time
  • You need full network access (npm install, API calls, database migrations)
  • You want parallel execution across multiple tasks
  • The local sandbox restrictions would block the work

For quick edits, file exploration, and interactive pair-programming, local execution is faster and more responsive.

The extension ships with integrated web search, enabled by default for local tasks 6. Two modes are available:

# ~/.codex/config.toml
web_search = "cached"  # Default — uses OpenAI's pre-indexed cache
# web_search = "live"  # Fetches live results (requires full-access sandbox)

The cached mode returns results from an OpenAI-maintained index rather than fetching live pages 6. This is faster and works within sandboxed sessions, but results may be hours or days behind. Live mode requires Agent (Full Access) or the equivalent sandbox_mode in config.toml.

Image Support

Drag and drop images into the prompt composer while holding Shift 6. VS Code normally intercepts drag events for its own file handling — the Shift modifier overrides this to pass the image to the Codex extension. This supports screenshots, design mockups, and error screenshots as context for prompts.

The TODO CodeLens Feature

When chatgpt.commentCodeLensEnabled is active, the extension adds CodeLens indicators above TODO comments in your code 8. Clicking the indicator triggers chatgpt.implementTodo, which sends the TODO and its surrounding context to Codex for implementation.

This creates a lightweight task-driven workflow: scatter TODOs as specifications, then let Codex resolve them one at a time with full file context.

VS Code as Multi-Agent Hub

Since February 2026, VS Code supports running multiple agentic extensions simultaneously 11. You can have Codex, Claude Code, and GitHub Copilot all active in the same editor, choosing between them per task. VS Code’s third-party agent API treats each extension as a separate agent with its own context and approval flow 12.

This multi-agent setup creates an interesting billing split: GitHub-hosted agents (Copilot) bill through your Copilot subscription, while provider extensions (Codex, Claude Code) bill through their respective subscriptions 12.

Marketplace Competition

As of early 2026, Claude Code leads Codex in VS Code Marketplace adoption — 5.2 million installs versus 4.9 million, with a higher average rating (4.0 vs 3.4 out of 5) 13. The competitive pressure has driven both teams to ship faster: Codex’s JetBrains integration, cloud delegation, and Spark model support all arrived in rapid succession.

Practical Configuration for Teams

A production team configuration combining the IDE extension with shared standards:

# .codex/config.toml (project-scoped)
model = "gpt-5.4"
approval_policy = "on-failure"
sandbox_mode = "workspace-write"
web_search = "cached"

[agents]
max_threads = 4
max_depth = 1

Combined with a well-structured AGENTS.md in the repository root, this gives every team member — whether they use the CLI, IDE extension, or cloud — the same agent behaviour with the same guardrails.

Limitations

Several CLI features do not yet work in the IDE extension:

  • Profiles — --profile flag is CLI-only; the IDE uses the top-level config only 7
  • Hooks — the experimental hooks engine fires in CLI sessions but not IDE sessions
  • Realtime/voice — the [realtime] audio pair-programming feature is CLI-only
  • Mid-turn steering — the CLI’s Enter-to-inject and Tab-to-queue features have no IDE equivalent
  • codex exec — non-interactive pipeline mode is a CLI concept; the IDE is inherently interactive

These gaps reflect the different interaction models: the CLI is a terminal-native tool with rich keyboard shortcuts, while the IDE extension integrates into VS Code’s panel-based UI.

Citations

  1. OpenAI Developers, “IDE extension – Codex”, 2026 — https://developers.openai.com/codex/ide ↩ ↩2 ↩3

  2. OpenAI Help Center, “Using Codex with your ChatGPT plan”, 2026 — https://help.openai.com/en/articles/11369540-using-codex-with-your-chatgpt-plan ↩

  3. JetBrains AI Blog, “Codex Is Now Integrated Into JetBrains IDEs”, January 2026 — https://blog.jetbrains.com/ai/2026/01/codex-in-jetbrains-ides/ ↩

  4. Neowin, “OpenAI’s Codex extension is now available for JetBrains IDEs”, January 2026 — https://www.neowin.net/news/openais-codex-extension-is-now-available-for-jetbrains-ides/ ↩ ↩2

  5. OpenAI Developers, “Models – Codex”, 2026 — https://developers.openai.com/codex/models ↩

  6. OpenAI Developers, “Features – Codex IDE”, 2026 — https://developers.openai.com/codex/ide/features ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  7. OpenAI Developers, “Configuration Reference – Codex”, 2026 — https://developers.openai.com/codex/config-reference ↩ ↩2 ↩3 ↩4

  8. OpenAI Developers, “Settings – Codex IDE”, 2026 — https://developers.openai.com/codex/ide/settings ↩ ↩2 ↩3

  9. OpenAI Developers, “Commands – Codex IDE”, 2026 — https://developers.openai.com/codex/ide/commands ↩

  10. LaoZhang AI Blog, “Claude Code vs Codex in 2026: Steer Live or Delegate Async?”, 2026 — https://blog.laozhang.ai/en/posts/claude-code-vs-codex ↩

  11. Visual Studio Code Blog, “Your Home for Multi-Agent Development”, February 2026 — https://code.visualstudio.com/blogs/2026/02/05/multi-agent-development ↩

  12. Visual Studio Code Docs, “Third-party agents in Visual Studio Code”, 2026 — https://code.visualstudio.com/docs/copilot/agents/third-party-agents ↩ ↩2

  13. Visual Studio Magazine, “Claude Code Edges OpenAI’s Codex in VS Code’s Agentic AI Marketplace Leaderboard”, February 2026 — https://visualstudiomagazine.com/articles/2026/02/26/claude-code-edges-openais-codex-in-vs-codes-agentic-ai-marketplace-leaderboard.aspx ↩