Codex vs Claude In-App Browsers: Six Workflows for an AI Operating System

Direct Answer

The important update is not that Codex and Claude now contain smaller versions of Chrome. It is that the browser is becoming a shared, task-scoped work surface. An agent can gather the pages needed for one outcome, act across approved sites, create drafts, and leave the result open for a person to inspect without breaking the thread's context.

Riley Brown demonstrates the pattern with YouTube analytics, Notion, X notifications, email, Google Docs, visual research, product comparisons, and podcast preparation. The strongest workflow is not autonomous posting. It is supervised preparation: the agent does the searching, tab opening, drafting, and field population; the human chooses, edits, sends, buys, or publishes.

Credits: The source video, demonstrations, four browser patterns, and daily workflow examples are by Riley Brown. Follow @rileybrown on X.

Source Note

Riley's video is the primary source for the July multi-tab demonstrations and his comparison of the two products. I checked the underlying capability and safety claims against current official documentation on 15 July 2026.

OpenAI's Browser documentation confirms a shared browser inside ChatGPT on web and desktop, manual navigation, browser comments, site allowlists and blocklists, direct sign-in, and Computer Use for clicking, typing, screenshots, and verification. It also says the built-in browser uses a separate profile from a person's regular browser and does not automatically inherit existing tabs or sessions.

Anthropic's Claude Code Desktop documentation confirms embedded app previews, automatic visual verification, annotation, parallel sessions, connectors, and Computer Use. Claude in Chrome remains Anthropic's documented route for browser tasks that need a signed-in Chrome session. Riley's arbitrary multi-tab browsing inside Claude Desktop appears newer than the currently indexed docs, so I treat that part as a creator test and rollout signal rather than a universal availability promise.

One mismatch is especially important. Riley shows Codex preparing X posts with local images, but OpenAI's current Browser page says the built-in browser cannot automate file uploads. That may reflect a different tool path, a staged rollout, or documentation lag. Test it on your own account and keep uploads behind review.

Link Map

What Actually Changed

The old agent-browser pattern often felt like a handoff to another machine. The agent launched a remote browser, opened Chrome, or returned a list of URLs. Context fragmented across the agent conversation, browser history, and whatever the user remembered to reopen.

The new pattern puts multiple browser tabs beside the task. In Riley's opening demonstration, one prompt asks Codex to open YouTube Studio analytics, a Notion video database, X notifications, important emails, a football score, and a generated support-ticket summary. Subagents gather the different surfaces while the parent task remains the organizing thread.

Three things become easier:

1. Gathering: ask for the workspace you need instead of manually assembling tabs.
2. Acting: let the agent prepare a document, draft, form, or application state inside the relevant tool.
3. Reviewing: keep each proposed result visible in its own tab so a person can compare, edit, approve, or reject it.

This is still not proof that every website, connector, login, upload, or action will work. Availability, account permissions, plan, operating system, site defenses, and rollout state all matter. The durable change is the interface model: the task owns the tabs.

Codex vs Claude: The Current Practical Difference

Do not convert a creator's personal setup into a permanent product ranking. Riley knows Codex better, has more of his context connected there, and says so. A Claude-heavy team may reach the opposite conclusion. Compare the exact workflow, not the logo.

Four Ways to Use the Browser

1. Open a tab manually

Start inside the task, open the browser, and navigate as a person would. The advantage is that the page is already beside the agent that knows the project. You can browse first, then ask the agent to summarize, edit, compare, or act on the visible state.

2. Ask the agent to open one known item

Use connected context to retrieve a specific document, record, or page: "Open the long-form video brief we were editing in Notion." This is retrieval with a visible handoff. The agent finds the item; the browser proves it opened the right one.

3. Ask the agent to assemble many tabs

This is the real upgrade. Ask for every video in progress, all relevant product options, each source behind a research conclusion, or the priority emails that need attention. Tabs become an evidence pack rather than a browsing accident.

4. Put the browser handoff inside a skill

Riley's email skill does more than draft replies. It verifies the current support threads, saves the replies, and opens each draft in its own tab with an explicit rule that nothing is sent. The repeatable behavior lives in the workflow, not in the user's memory.

Six Practical Browser Workflows

1. Email triage with unsent drafts

Ask the agent to identify important threads, read the full conversation, prepare a reply, and open every draft separately. Keep sending manual until the workflow has a strong test history. This gives the agent the repetitive preparation work without handing it your reputation.

2. Visual B-roll research in script order

Riley gives Codex a video script and asks it to open image searches in the order each visual would appear. Instead of returning a text list, it creates a tab sequence the editor can review line by line. A better production version would also record the source URL, usage rights, intended timestamp, and approval status.

3. Social post preparation with assets

The agent can analyze images, propose captions, and prepare multiple post drafts. The useful phrase in Riley's earlier X example is "do not actually tweet it." Publishing remains the human gate. As noted above, current OpenAI docs do not promise automated uploads in the built-in browser, so treat local-image attachment as rollout-dependent.

4. Decide the next task, then prepare the workspace

Riley asks the agent to review calendar, email, Notion, and deadlines, then offer the three most important next tasks. After he chooses podcast preparation, the agent opens guest documents, profiles, previous interviews, and relevant research, then helps draft the introduction in the correct document.

This is more powerful than a generic daily brief because the output is not only advice. It is a ready-to-work environment. The risk is letting the model decide importance without a clear objective, so provide deadlines, revenue or customer impact, dependencies, and exclusions.

5. Product research with inspectable evidence

One of Riley's parallel sessions researches a Windows PC for multi-camera filming with vMix. The agent creates a summary document and opens the relevant product pages. That makes it easier to challenge recommendations, compare real specifications, and catch stale prices or incompatible hardware.

6. Parallel content and guest research

Riley launches separate sessions for YouTube hook analysis and podcast guest sourcing while the hardware research runs. Each session produces a document plus source tabs, including direct videos and timestamps where useful. Parallelism is valuable only when each thread has a distinct question and output; otherwise it creates three noisy research piles instead of one decision.

Why This Feels Like an Operating System

An operating system coordinates resources around work. The emerging AI workspace coordinates model context, browser tabs, local files, apps, connectors, skills, subagents, permissions, and human review around an outcome.

Riley borrows a useful idea from Jack Dorsey: shift from telling agents what to do toward asking what should be done, then pull the best thread. That only works when the agent can inspect enough authorized context and when the person remains responsible for the decision.

The browser matters because most business software already lives there. It is the compatibility layer for email, CRM, analytics, content, documents, support systems, and internal tools. The agent does not need a perfect API for every edge case if it can use a controlled browser for the last mile.

But the operating-system analogy can encourage dangerous overreach. A real operating model needs separation:

• Thread: one objective and its conversation history.
• Tabs: the pages and review states needed for that objective.
• Connectors: structured access to approved business data and actions.
• Skill: repeatable instructions, checks, and handoff rules.
• Computer Use: a broader fallback for visual interfaces that structured tools cannot reach.
• Human gate: the irreversible or reputational decision.

The Permission Model Matters More Than the Tabs

A browser page is untrusted input. It may contain outdated information, misleading instructions, hostile prompt injection, or a destructive button beside the correct one. OpenAI explicitly warns users to treat page content as untrusted and asks for confirmation before sensitive actions such as purchases, permission changes, submissions, or deletion.

Anthropic similarly documents Computer Use as broader and slower than connectors, shell tools, or Chrome. It is off by default, works on the actual desktop, and requires app permissions. That ordering is sensible for both platforms:

1. Use a read-only connector or API when one exists.
2. Use the browser when visible page state matters.
3. Use Computer Use only when a graphical interface is unavoidable.
4. Require a person for consequential actions.

Keep separate browser profiles or dedicated work accounts where practical. Do not expose personal banking, password managers, private messages, admin consoles, or unrelated tabs merely because the agent can browse. Narrow access is a product feature, not an inconvenience.

Copy-Ready Browser Agent Policy

Browser workflow policy

Goal
- Complete: [one specific outcome].

Allowed sources
- Read only: [domains, folders, accounts, or connectors].
- Write drafts only in: [approved tools].

Allowed actions
- Open tabs, search, summarize, compare, and populate draft fields.
- Create a review document with source links and timestamps.

Never do without approval
- Send email or messages.
- Publish or schedule content.
- Upload files.
- Submit forms containing personal or company data.
- Buy, subscribe, accept terms, or change permissions.
- Delete, archive, merge, or overwrite records.

Verification
- Keep every material source open or linked.
- Mark uncertain or conflicting information.
- Confirm the target account before preparing a write action.
- Stop if a page asks for a secret, CAPTCHA, identity check, or new permission.

Handoff
- Open each draft or recommendation in a separate labeled tab.
- State what changed, what remains unverified, and the exact action awaiting review.

Adapt the policy to the risk. A public-source research task can be broad and read-only. A support inbox needs approved accounts, customer-data rules, a tone guide, escalation criteria, and a strict no-send boundary.

A Seven-Day Adoption Plan

Bottom Line

Riley Brown is right about the direction. In-app browser tabs make Codex and Claude feel less like chatbots beside the work and more like workspaces that can assemble, operate, and present the work.

The winning pattern is not maximum autonomy. It is minimum friction between context, action, and review. Ask the agent to gather the right surfaces, prepare the reversible parts, preserve the evidence, and stop at the human decision.

Codex currently looks more complete in Riley's setup for knowledge-work tab orchestration. Claude's desktop, Chrome, preview, annotation, and Computer Use surfaces are moving toward the same destination. Product details will change quickly. The durable skill is designing a browser workflow that remains useful when the model or app changes: one outcome, narrow access, visible evidence, a clear stop condition, and a human gate.

Sources

• Riley Brown: Codex and Claude Shipped Browser Updates. This Changes Everything.
• Riley Brown on YouTube and Riley Brown on X
• OpenAI: Browser
• OpenAI: Computer Use
• OpenAI: ChatGPT for your most ambitious work
• Anthropic: Claude Code Desktop
• Anthropic: Use Claude Code with Chrome
• Anthropic: Claude Code platforms and integrations