Codex Chrome Extension + Codex App: A Working Guide for the May 2026 Launch

TL;DR

On May 7, 2026, OpenAI released the Codex Chrome extension. The significant advancement is that the agent now operates with your actual Chrome profile, enabling access to authenticated web applications like Salesforce, Gmail, and internal dashboards without requiring separate credential management. Site-based permissions operate on a per-request confirmation basis by default. The initial rollout experienced some challenges, and EU/UK regions are currently excluded from access.

The extension represents browser automation capabilities enhanced by existing session authentication rather than entirely new functionality.

If you currently rely on Codex CLI for daily development tasks, the related "Codex CLI real-world coding workflow" article covers terminal-based usage patterns. This guide addresses the Chrome integration: its capabilities, limitations, and role within a contemporary 2026 development environment.

What actually launched on May 7

OpenAI released two interconnected products during early May. The Chrome extension became available on May 7, 2026, concurrent with the Codex desktop application's Plugins system launch. The official Codex changelog states:

"With the new extension for Chrome, Codex is even better at working with apps and websites in your browser. It works in parallel across tabs in the background without taking over your browser, and you stay in control of which websites Codex can use."

One week afterward, on May 14, OpenAI introduced remote connectivity capabilities—permitting Codex operation from the ChatGPT mobile application while connected to a Mac running the Codex desktop application. Identical projects, plugins, and Chrome sessions remain accessible, simply managed through your mobile device. These concurrent updates function as complementary pieces: the Codex application serves as the central hub, while Chrome and mobile interfaces become the interaction surfaces.

OpenAI reported significant growth metrics at launch: Codex now maintains more than 4 million weekly active users, representing an 8x expansion since January. Regardless of definitional nuances surrounding this metric, the trajectory demonstrates clear commitment. Codex represents OpenAI's focused development investment for the developer community.

Install: the four steps that actually work

Though the initial week generated considerable confusion (addressed subsequently), the functional installation process remains concise:

1. Launch the Codex desktop application and navigate to the Plugins menu
2. Select Add for the Chrome plugin
3. Follow the installation instructions—this directs you to the Chrome Web Store Codex extension page and installs version 1.1.4
4. Authorize Chrome's permission requests, then confirm the extension displays Connected within Chrome's toolbar

The non-obvious requirement involves step one: the extension installation occurs through the Codex desktop application, not the Chrome Web Store independently. Though the extension exists on the store, functionality requires the desktop application handshake. Several early-week developers attempted the reverse sequence and received a non-functional extension.

Should the Chrome plugin remain absent from your Plugins menu, you likely belong to an early-access cohort awaiting activation. Update the Codex desktop application to the latest version, sign out and back in, and verify again. Regional limitations also govern plugin availability—refer to the EU/UK section below.

What the extension can actually do

The promotional description "Codex works in your browser" represents accuracy without sufficient specificity. The concrete capabilities include:

• Drive sites where you are signed in. The primary use case. Gmail, Salesforce, LinkedIn, Notion, enterprise SSO-protected applications, and any platform your standard Chrome profile can access.
• Read across multiple tabs. Request Codex to compare product pages, reconcile specification documents with tickets, or extract context from Linear boards while developing code addressing the ticket.
• Use Chrome DevTools. The extension provides the page debugger, allowing Codex to examine failing components, review console messages, and monitor network communications instead of relying on assumptions.
• Run in parallel without stealing focus. The background-tab design approach makes this genuinely usable. Previous browser agents would commandeer the foreground window repeatedly. This version operates without interference.
• Reference your browsing history. Scoped to individual requests with confirmation requirements. Helpful for "locate that article I read last Tuesday regarding X" scenarios, less valuable for broader applications.

The capability delivering substantial practical benefit operates quietly: multi-tab context. A significant percentage of authentic coding assignments involve "examine this Stripe documentation, this internal endpoint, and this GitHub pull request simultaneously, then complete the integration." Previously, extension unavailability forced you to copy all three into your message. Currently, Codex accesses them directly.

The permissions model — what you are actually agreeing to

The Chrome Web Store listing requires these permissions:

• Access the page debugger
• Read and change all your data on all websites
• Read and change your browsing history on all your signed-in devices
• Display notifications, manage bookmarks, handle downloads

This represents a substantial scope. Two mitigations exist:

1. Per-site confirmation operates by default. Each new domain Codex accesses generates a confirmation request. Approval grants permission for "this specific site, this particular task." Selecting always-allow disables the prompt for that domain permanently.
2. Allowlist and blocklist apply layered restrictions. Pre-approve a limited number of sites and block others completely. Codex respects these limitations regardless of underlying technical capabilities.

The recommended configuration for a developer using this on professional hardware:

• Keep per-site confirmation active
• Permanently allow the one to three sites you are actively automating (your CRM, your monitoring interface, your documentation)
• Block anything involving financial data, HR functions, and anything where errors create significant risk

Treat the extension as you would granting a junior developer access to your laptop. Appropriate for clearly defined assignments, inappropriate for "determine what needs doing."

Where it fits: three real use cases

Practical scenarios where this has demonstrated genuine value during the initial week:

1. CRM and inbox triage. "Examine the three Salesforce opportunities marked this week, retrieve the final email correspondence from each, and prepare a response referencing current call details." Without the extension, you would handle this manually or implement Salesforce and Gmail API integrations. With the extension, Codex operates the actual web interface. Slower per operation, yet requires no authentication configuration.

2. Web app debugging with real session state. A bug appearing exclusively when signed in as a particular user position. The extension can navigate to the staging platform, authenticate (you maintain existing authentication), proceed through the broken workflow, access console messages using DevTools, and suggest corrections in your local codebase. The headless-browser equivalent has consistently remained technically feasible yet practically challenging to establish.

3. Cross-tab research. "Read this OpenAI article, this Anthropic document, and this benchmark resource, then prepare the transition strategy." Codex examines three tabs in-place. The response reflects genuine current content from each resource, not model training information.

For terminal-only code operations—refactoring, document management, command execution—continue using Codex CLI. The Chrome extension represents the incorrect selection for these functions. For comprehensive agent comparison, the "coding agents head-to-head" article indicates where each tool performs optimally.

Where it fits in an ofox-based stack

Direct context: the Chrome extension connects exclusively to the Codex desktop application, which validates through your OpenAI credentials. It presently lacks support for OPENAI_BASE_URL redirection, preventing Chrome extension traffic routing via aggregators such as ofox.ai. This constitutes an OpenAI-direct billing relationship.

Available alternatives involve dividing your tools:

• Codex Chrome + Codex desktop application → OpenAI direct, for authenticated browser activities
• Codex CLI → directed via ofox with OPENAI_BASE_URL=https://api.ofox.ai/v1 and your ofox credentials, for headless terminal operations

This separation is not a compromise approach. It represents the most efficient method of conceptualizing the workflow. The extension's advantage centers on Chrome profile access, and routing infrastructure cannot address this. CLI capabilities emphasize portable, automatable code editing, where routing infrastructure delivers GPT-5.3 Codex pricing at $1.75 / $14 per million input/output tokens via ofox, including the capacity to test Claude or Gemini alternatives through identical credentials. The "AI API aggregation guide" explains the routing methodology.

For routing configuration, the "Codex CLI configuration guide" outlines required environment settings. The "Codex installation guide" addresses desktop application setup if necessary.

What is still broken in week one

Initial week of significant deployment. Important considerations:

• Plugin missing from directory. The most frequent complaint—the Chrome plugin fails to appear within the Codex application's Plugins menu. This involves phased deployment. If your account hasn't activated yet, update the application and re-authenticate to occasionally enable access. Otherwise, expect OpenAI to increase rollout progressively.
• Extension install fails with a generic error. Typically indicates outdated Codex application version. Refresh the desktop application to current release before retrying.
• EU and UK are excluded. OpenAI has confirmed extension unavailability in these locations at launch and hasn't specified timeline for future release. Codex CLI and non-Chrome desktop functions continue functioning; the extension specifically remains restricted.
• No Chromium-based browser support yet. Edge, Brave, Arc, Opera remain unsupported despite Chromium foundation. GitHub issue #22638 documents this limitation. OpenAI recognizes but hasn't specified implementation timeline.
• Performance varies with tab count. Background operation represents the design philosophy, though Chrome profiles maintaining 80+ tabs will measurably reduce Codex performance. For users with extensive tab collections, operate the extension in a separate Chrome profile.

Where this goes next

The Chrome extension combined with the May 14 remote connectivity announcement suggest a two-phase trajectory. Initial phase: establish Codex's ability to operate your authentic browser. Secondary phase: enable Codex remote access via your mobile device. Combined, the system represents a desktop-anchored agent accessible from anywhere while retaining your login authentication. This constitutes substantially different functionality from "an LLM within a messaging interface."

For development teams, the practical approach for the upcoming quarter involves maintaining code workflows via CLI (directed via your preferred provider) while progressively transitioning web-application tasks, incrementally per site, toward the Chrome extension. CRM modifications, support management, internal-system automation. Begin with the allowlist restricted to one domain. Introduce additional sites only after the initial site demonstrates reliable performance.

The optimal approach for integrating the Codex Chrome extension mirrors integrating a junior engineer: permitting one site incrementally, with confirmation prompts active, and limiting potential impact to non-critical systems.

For examining how the Codex Chrome extension compares with competing solutions within a comprehensive development setup, the "best LLM for coding ranked by real use" article addresses model evaluation, and the "coding agents head-to-head" piece covers tooling comparison.

---

Originally published on ofox.ai/blog.