Atlas Sunsets on 9 August: Migrating Browser-Agent Workflows to Codex CLI, Playwright MCP, and the ChatGPT Desktop App

OpenAI’s Atlas browser — launched on 21 October 2025 as a standalone desktop browser with AI at its core — stops working on 9 August 2026 ¹. That gives teams eight days to migrate any browser-dependent agentic workflows before the lights go off. For developers who had woven Atlas into their Codex CLI pipelines — previewing localhost builds, automating form-heavy test scenarios, or orchestrating cross-site research — the question is not whether to move, but where each workflow lands in the post-Atlas landscape.

This article maps every Atlas capability to its successor surface and shows how to configure Codex CLI for the transition.

What Atlas Was — and Why It Mattered to Developers

Atlas shipped as a Chromium-based desktop browser with ChatGPT embedded directly in the navigation experience ¹. Its Agent mode, previewed in early 2026, could autonomously visit pages, fill forms, download files, and maintain authenticated sessions across sites like GitHub, Linear, Jira, and Figma — all without the copy-paste friction of switching between a browser and an AI assistant ².

For Codex CLI users, Atlas offered something no MCP server could: a visual browser with persistent login state running on the same machine as the agent. Developers used it to:

Preview localhost renders during front-end iteration loops
Verify deployed staging environments mid-task
Scrape authenticated internal tools (wikis, dashboards, admin panels)
Drive form-heavy QA scenarios that Playwright scripts had not yet covered

The nine-month experiment is ending because OpenAI is consolidating Atlas’s capabilities into two durable surfaces: the ChatGPT desktop application and ChatGPT Work ².

The Replacement Architecture

The post-Atlas world splits browser-agent capabilities across three surfaces, each with a distinct scope:

graph TD
    A[Atlas Browser<br/>Sunset 9 August] --> B[ChatGPT Desktop App<br/>Built-in Browser]
    A --> C[ChatGPT Work<br/>Cloud Browser Agent]
    A --> D[Codex CLI +<br/>Playwright MCP Server]

    B -->|Authenticated browsing<br/>Multi-tab navigation<br/>Computer Use| E[Visual Workflows]
    C -->|Research & analysis<br/>Document creation<br/>Public URL agent tasks| F[Cloud Agent Tasks]
    D -->|Localhost preview<br/>Accessibility snapshots<br/>Form interaction<br/>Screenshot verification| G[Developer Workflows]

    style A fill:#ff6b6b,color:#fff
    style B fill:#4ecdc4,color:#fff
    style C fill:#45b7d1,color:#fff
    style D fill:#96ceb4,color:#fff

ChatGPT Desktop App: Authenticated Browsing and Computer Use

The merged ChatGPT desktop application — which now houses Chat, Work, and Codex on every plan, including Free ³ — inherits Atlas’s built-in browser with support for multiple tabs, downloads, improved navigation, and account login persistence ². Computer Use, powered by GPT-5.6, handles click-type-move operations for one-off or scheduled tasks ³.

For developers, this covers the “visual verification” use case: reviewing rendered pages, annotating UI issues, and instructing the agent to fix them without leaving the window ⁴.

ChatGPT Work: Cloud-Hosted Agent Tasks

ChatGPT Work, launched on 9 July 2026, handles research, analysis, and document creation using a cloud-hosted browser ³. It breaks goals into steps and returns finished artefacts — spreadsheets, presentations, reports, and Sites.

The critical architectural difference from Atlas: Work’s browser runs in the cloud, not locally ⁵. Workflows that required access to localhost or internal network services cannot migrate here without rearchitecting.

Codex CLI + Playwright MCP: The Developer Path

For Codex CLI users, the most capable replacement for Atlas’s developer-facing workflows is the Playwright MCP server ⁶. Unlike Atlas’s visual browser, Playwright MCP operates through structured accessibility snapshots — bypassing the need for screenshots or visually-tuned models — and provides deterministic, scriptable browser interaction.

Configuring Playwright MCP for Codex CLI

The Playwright MCP server integrates with Codex CLI through the standard MCP configuration in config.toml:

[mcp_servers.playwright]
command = "npx"
args = ["@anthropic-ai/mcp-playwright"]
enabled = true

Alternatively, register it via the CLI:

codex mcp add playwright -- npx @anthropic-ai/mcp-playwright

The server exposes tools for navigation, element interaction, screenshot capture, and accessibility tree inspection ⁶. For localhost preview workflows — the most common Atlas use case among Codex CLI developers — this provides a direct replacement:

# Start your dev server, then use Codex CLI with Playwright MCP
npm run dev &
codex "Navigate to http://localhost:3000, verify the login page renders correctly, \
       check that the form fields are accessible, and screenshot the result"

Scoping Playwright Access in the Sandbox

Codex CLI’s sandbox restricts network access by default. For Playwright MCP to reach localhost or staging URLs, configure the network proxy allowlist:

[sandbox]
network_proxy_allowed_domains = [
  "localhost",
  "127.0.0.1",
  "*.staging.internal.example.com"
]

For production-like environments behind authentication, pair Playwright MCP with a cookie-injection pattern or use environment variables for service tokens:

[env]
STAGING_TOKEN = "env:STAGING_AUTH_TOKEN"

Migration Decision Matrix

Not every Atlas workflow maps to the same successor. Use this decision matrix:

Atlas Workflow	Successor	Configuration
Localhost front-end preview	Codex CLI + Playwright MCP	`config.toml` MCP server + sandbox allowlist
Authenticated site browsing (GitHub, Jira, Figma)	ChatGPT Desktop App browser	Built-in, no CLI configuration needed
Autonomous web research	ChatGPT Work	Cloud-hosted, no CLI configuration needed
Form-heavy QA automation	Codex CLI + Playwright MCP	Accessibility snapshots + element interaction tools
Visual regression checking	Codex CLI + Playwright MCP	Screenshot tool + comparison in AGENTS.md workflow
Cross-site data gathering	ChatGPT Work or Codex CLI web search	`codex --web-search` or indexed search mode
Computer Use (click-type-move)	ChatGPT Desktop App	GPT-5.6 powered, desktop-only

AGENTS.md Patterns for Browser-Agent Workflows

With Atlas gone, browser interaction via Playwright MCP requires explicit guidance in your AGENTS.md to prevent the agent from attempting browser operations without the MCP server:

## Browser Interaction Rules

- ALWAYS use the Playwright MCP server for browser interaction — never attempt
  to launch a browser directly
- For localhost previews, ensure the dev server is running before navigation
- Use accessibility snapshots over screenshots where possible — they consume
  fewer tokens and provide structured data
- Screenshot only for visual regression verification or layout debugging
- Never store authentication credentials in AGENTS.md — use environment
  variables via config.toml [env] section

Data Export Before 9 August

Atlas data does not transfer automatically ². Before the shutdown:

Bookmarks: Export to Chrome or the ChatGPT desktop app
Passwords and cookies: Export to the ChatGPT desktop app’s integrated browser
Browser history: No automated export — manually save any critical URLs
Saved pages or downloads: Already on your local filesystem; no action needed

For teams that built Atlas-dependent automations (scheduled browsing tasks, recurring form submissions), audit these now and rebuild them as either ChatGPT Work tasks or Codex CLI + Playwright MCP workflows.

The Architectural Lesson

Atlas’s nine-month lifespan underscores a pattern that senior developers should internalise: OpenAI’s durable integration surfaces are the ChatGPT API, Codex (both CLI and cloud), and the Realtime API ⁵. Standalone product experiments like Atlas are useful for proving capabilities but carry deprecation risk.

For Codex CLI users, the lesson is practical: build browser-dependent workflows on the MCP server layer, not on first-party GUI products. Playwright MCP, BrowserBase MCP, and Puppeteer MCP servers all provide browser automation through the stable MCP protocol — a protocol that survives product consolidation because it sits below the application layer.

graph LR
    subgraph Stable["Stable Integration Layer"]
        A[Codex CLI] --> B[MCP Protocol]
        B --> C[Playwright MCP]
        B --> D[BrowserBase MCP]
        B --> E[Puppeteer MCP]
    end

    subgraph Volatile["Volatile Product Layer"]
        F[Atlas Browser<br/>Oct 2025 – Aug 2026]
        G[Future Product X?]
    end

    style F fill:#ff6b6b,color:#fff,stroke-dasharray: 5 5
    style G fill:#ffcc00,color:#333,stroke-dasharray: 5 5
    style Stable fill:#e8f5e9,stroke:#4caf50
    style Volatile fill:#ffebee,stroke:#f44336

What Changes on 10 August

After 9 August 2026:

Atlas will no longer open or support browser-based agentic workflows ²
The ChatGPT desktop app’s built-in browser becomes the primary visual browsing surface
ChatGPT Work handles cloud-hosted autonomous browser tasks
Codex CLI + Playwright MCP remains the developer-native path for programmatic browser interaction
No Codex CLI configuration changes are required — the CLI never depended on Atlas directly

The transition is less a migration and more a clarification: Atlas was always a preview of capabilities that now live in more durable homes. For CLI-first developers, nothing breaks on 10 August — provided you were already using MCP servers rather than relying on Atlas as a manual bridge.

Citations

OpenAI Pulls Plug on ChatGPT Atlas Browser After Nine Months — WebProNews, July 2026 ↩ ↩²
Evolving Atlas into ChatGPT for browser-based agentic work — OpenAI Help Centre, July 2026 ↩ ↩² ↩³ ↩⁴ ↩⁵
ChatGPT is now a partner for your most ambitious work — OpenAI, July 2026 ↩ ↩² ↩³
Codex Merged With ChatGPT App: What Changed & How to Access — Coursiv, July 2026 ↩
OpenAI Atlas Browser Shuts Down August 9, 2026: Migration Guide — ChatForest Builder’s Guide, July 2026 ↩ ↩²
Playwright MCP Server Documentation — Microsoft Playwright, 2026 ↩ ↩²