How to Use Codex for Research, Projects, and Business Work

OpenAI Codex is no longer limited to writing and fixing code. The current Codex experience can work with project context, local files, repositories, browser-based information, plugins, connected business tools, and longer-running goals. OpenAI also now positions Codex as a tool for broader knowledge work, including research, data analysis, reports, and business workflows.

There is one important distinction before you begin: ChatGPT Work is the better choice when your main objective is general research and finished reports, while Codex is designed primarily for technical work and agentic workflows involving code, files, tools, and connected systems. On the current desktop app, ChatGPT, Work, and Codex are separate experiences.

This guide shows where Codex fits, how to set it up, and how to use it effectively for research, ongoing projects, and business work.

What You Need

Essential

• A ChatGPT account with Codex access
• The current ChatGPT desktop app for Windows or macOS when you want to use the full desktop Codex experience
• Access to the files, repositories, websites, or connected apps required for your task
• Permission to access any external accounts or workspace resources you connect

Codex is currently included across ChatGPT plans, although usage limits vary by plan. The desktop app combines Chat, Work, and Codex, while Codex itself remains a separate experience inside the desktop application.

Optional

You may also benefit from:

• Existing project files or reference documents
• A Git repository
• Connected apps
• Codex plugins
• A business or team workspace
• A clearly defined goal and success criteria

Difficulty and Time

Difficulty: Easy to Moderate

Initial setup: About 5–15 minutes

Typical task time: Varies considerably depending on the amount of research, files, connected tools, and work Codex needs to perform.

Before You Start

Think about the result you actually want before opening Codex.

Instead of asking:

"Research AI trends."

Give Codex a measurable objective such as:

"Research the latest AI agent developments, compare five major products, identify the most important changes from the last 12 months, and produce a source-backed report with a comparison table and recommendations."

This matters because Codex is designed to work toward goals and can continue through longer, multi-step tasks. Goal mode is available across the Codex app, IDE extension, and CLI.

Install or Update the ChatGPT Desktop App

The current desktop experience combines Chat, Work, and Codex on Windows and macOS. If you are still using the older Codex desktop application, OpenAI says updating it converts it into the new ChatGPT desktop app while retaining existing Codex chats and projects.

1. Download the current ChatGPT desktop application from OpenAI.
2. Install it on Windows or macOS.
3. Open the application.
4. Sign in with the ChatGPT account you want to use with Codex.

OpenAI's current desktop documentation directs users to the official ChatGPT download page for the latest application.

Open Codex

Once the desktop app is running:

1. Look at the top-left of the ChatGPT desktop application.
2. Select Codex.
3. Start a new Codex thread.

On the current desktop interface, Codex remains a separate experience from ChatGPT and Work. Its history is also kept separate from normal ChatGPT history.

Codex can work with local folders, repositories, terminals, developer tools, browser context, and other capabilities depending on your setup and permissions.

Give Codex a Clear Goal

A strong Codex task starts with a specific outcome.

Instead of:

Research my project.
Goal:
Research the current market for AI-powered customer-support software.
Success criteria:
- Compare at least five major products.
- Use current, reputable sources.
- Separate verified facts from estimates or opinions.
- Identify pricing, major features, target customers, and important limitations.
- Produce a concise report with a comparison table.
- Include the source for each important factual claim.

Use:

This gives Codex both the task and the definition of success.

Codex's Goal mode is designed around exactly this type of outcome-oriented work.

Add the Right Context

Codex becomes more useful when it has access to the material it actually needs.

Depending on the task, provide:

• Relevant documents

• Spreadsheets

• Project files

• Existing reports

• A code repository

• Approved connected apps

• Web pages or browser context

• Team instructions

• Reusable skills or plugins

Do not dump unrelated material into the task. More context is not automatically better. The goal is to give Codex the information necessary to make a reliable decision or produce the requested deliverable.

For broader ChatGPT Projects, OpenAI currently supports files, project instructions, saved responses, supported app links, and connected apps as project context. Projects are intended for long-running work such as research, planning, reporting, and content creation.

Use Codex for Research When the Research Is Part of a Larger Workflow

Codex can now support research and knowledge work, especially when research leads directly to another action.

For example, you can ask it to:

• Examine information from approved sources
• Collect and compare data
• Analyze files
• Investigate an issue
• Turn findings into a report
• Create a dashboard or lightweight tool
• Update an existing workflow based on the findings

OpenAI says teams are using Codex for research, data analysis, reports, presentations, spreadsheets, and other knowledge-work tasks.

A practical research prompt looks like this:

A practical research prompt looks like this:

Investigate the current state of AI coding agents.
Research objectives:
1. Identify the five most significant developments from the past 12 months.
2. Prioritize official company documentation and reputable technical sources.
3. Separate confirmed product capabilities from announced or experimental features.
4. Compare the major tools on coding, research, automation, integrations, and pricing.
5. Produce a source-backed report.
6. Flag anything that could not be independently verified.

Important distinction: Codex or ChatGPT Work?

For pure research, OpenAI currently recommends the Work experience rather than treating Codex as a replacement for a dedicated research workflow.

OpenAI describes Work as the experience for researching topics, analyzing information, and producing documents, spreadsheets, presentations, reports, and Sites. Codex is described as the experience for writing or debugging code, running tests and commands, reviewing changes, and working with repositories.

That means the best workflow is often:

Work → research and analyze

Codex → execute technical or tool-based work resulting from that research

For example, use Work to investigate competitors, then use Codex to build a comparison dashboard from the resulting data

Use the Built-In Browser When a Task Needs a Web Page

The current ChatGPT desktop application includes a built-in browser that can be used from both Work and Codex.

To use it:

1. Open a Work or Codex chat.
2. Open the browser from the toolbar.
3. On macOS, the documented shortcut is Command + Shift + B.
4. On Windows, use Ctrl + Shift + B.
5. Open or provide the relevant webpage.
6. Approve access when prompted.
7. Ask Codex to inspect or work with the page.

OpenAI says the browser can browse pages, work across tabs, download files, and handle supported sign-in workflows. Availability can depend on your plan and workspace settings.

Example

Open the official documentation for this product.
Compare the current setup requirements with our existing implementation.
Identify anything that has changed.
Do not change files yet.
First give me a short list of required updates and cite the relevant documentation.

This is safer than telling an agent to immediately make changes based on unverified information.

Organize Long-Running Work in a Project

Projects are useful when the task will continue over multiple conversations.

For example, create a project for:

Website Redesign

Then keep related material there:

• Design requirements
• Existing content
• Brand instructions
• Research
• Client feedback
• Planning documents
• Previous decisions
• Relevant files

OpenAI describes Projects as workspaces for keeping chats, files, instructions, and ongoing context together. They can also be reused for recurring workflows.

Create a ChatGPT Project

1. Open ChatGPT.
2. Select New project in the sidebar.
3. Give the project a clear name.
4. Add the relevant files.
5. Open Project settings.
6. Add project-specific instructions.
7. Add useful sources from supported apps, conversations, or pasted reference material.

Projects can also use connected apps, including supported sources such as Google Drive and Slack.

Keep Project Instructions Specific

Project instructions should tell the AI how the work should be handled.

For example:

Project instructions:
You are assisting with an ongoing technology website project.
Priorities:
- Accuracy over speed.
- Use current sources.
- Do not invent technical specifications.
- Keep existing brand terminology consistent.
- Flag uncertain information.
- Preserve previously approved decisions unless I explicitly change them.
- Before making major changes, explain what will be affected.

Project instructions apply within that project and are designed to keep responses consistent across ongoing work.

Use Saved Research as Reusable Context

One useful Project workflow is saving valuable previous outputs.

For example:

1. Complete a market-research task.
2. Review the resulting analysis.
3. Save the useful response to the Project.
4. Reuse that information in later work.

OpenAI currently supports saving ChatGPT responses to Project sources so useful analysis, decisions, drafts, or summaries can be reused later.

This is particularly useful for large projects where the same research will be referenced repeatedly.

Use Codex Plugins for Business Work

Codex can be extended through plugins that package skills, connected apps, and reusable workflows.

To use a plugin in Codex:

1. Open the Sources area in a supported Codex task view.
2. Select Use plugins.
3. Search for the plugin you need.
4. Select the installed plugin.
5. Complete any required setup or account authorization.
6. Give Codex a task that clearly explains what you want the connected tool to do.

OpenAI says plugins can package reusable instructions, connected apps, and workflow capabilities. The exact plugins available depend on the account, workspace, region, plan, and surface.

Choose the Business Workflow That Matches the Job

Codex can now support a range of business workflows beyond software development.

OpenAI's current role-specific Codex plugins include workflows for areas such as:

• Data analytics
• Sales
• Product design
• Creative production
• Public-equity investing
• Investment banking

These workflows can connect Codex to business tools and help turn information into reports, analysis, follow-ups, dashboards, or other deliverables.

Example: Business analysis

Analyze this quarter's sales data.

Tasks:

- Compare this quarter with the previous quarter.
- Identify the largest changes.
- Highlight unusually strong or weak segments.
- Explain likely drivers only when supported by the available data.
- Produce a management-ready summary.
- Clearly separate observed facts from interpretation.

The important part is that Codex should receive the actual business context instead of being asked to guess what the company is trying to achieve.

Connect Business Apps Only When Necessary

A plugin may use connected apps to retrieve information or perform supported actions.

The app still operates within the permissions granted to that account or workspace. Installing a plugin does not bypass provider permissions or workspace controls.

For example, a workflow may require access to:

• Slack
• Google Drive
• GitHub
• GitLab
• CRM data
• Analytics platforms
• Internal business systems

Connect only the services the task genuinely needs.

For a business workspace, administrators can control plugin and app access, installation policies, permissions, and related workspace settings.

Keep Human Approval in the Workflow

Do not treat Codex's first result as automatically correct.

A strong business workflow is:

Research → Analyze → Draft → Review → Approve → Execute

For example, instead of:

Update everything you find.

use:

Analyze the available customer data and prepare the recommended changes.

Do not make external changes yet.

Show me:

1. What you found.
2. What you recommend changing.
3. Which data supports each recommendation.
4. What could go wrong.

Wait for approval before taking external actions.

This is especially important when an action could alter customer information, business records, production systems, or other shared resources.

OpenAI's business-oriented Codex guidance emphasizes keeping people in control of review and approval.

Use Codex to Turn Research Into a Deliverable

One of the strongest workflows is connecting research directly to an output.

For example:

Research the current competitive landscape for AI website builders.

Then:
1. Compare the major competitors.
2. Identify their strongest differentiators.
3. Create a structured competitive-analysis report.
4. Build a spreadsheet containing the comparison.
5. Create a simple dashboard showing the most important metrics.
6. Clearly label assumptions and unverifiable claims.
7. Do not publish or send anything externally.

OpenAI says Codex is increasingly being used to turn research and information into reports, spreadsheets, presentations, dashboards, and lightweight tools.

For Business and Enterprise customers, Codex can also create and share interactive hosted websites and apps through Sites, which OpenAI introduced as a preview for those customers.

Use Multiple Agents for Larger Projects

Codex is designed to handle parallel work.

For a large project, separate the work into independent tasks rather than giving the entire project one enormous instruction.

For example:

Agent 1: Research competitors

Agent 2: Analyze customer data

Agent 3: Review technical requirements

Agent 4: Build the reporting dashboard

Agent 5: Review the final output

The Codex app is built around coordinating multiple agents and running work in parallel.

Parallel work is most useful when the tasks are genuinely independent. Splitting one tightly connected task into too many agents can create unnecessary coordination.

Use a Review Prompt Before Accepting the Result

After Codex finishes, ask it to review its own work.

A useful review prompt is:

Review your completed work before I approve it.

Check:

- Factual accuracy
- Missing requirements
- Unsupported assumptions
- Broken references
- Incorrect calculations
- Security or privacy risks
- Incomplete steps
- Unexpected changes
- Any claim that needs additional verification

Do not change anything yet.

First give me the review findings.

For technical tasks, Codex can also work with tests, linters, type checkers, repositories, and other developer tools.

Verify the Sources

Research quality depends heavily on source quality.

For important claims, ask Codex to distinguish between:

Verified: Supported by an authoritative or reliable source.

Reported: Mentioned by a reputable secondary source but not independently confirmed.

Inferred: A conclusion based on available evidence.

Unknown: Not sufficiently supported.

This approach is particularly useful for business research, product comparisons, market intelligence, pricing research, and technology reporting.

Do not accept a polished answer as proof that the underlying facts are correct.

Protect Sensitive Business Information

Business workflows can involve confidential data, so use the smallest amount of information needed for the task.

Before connecting an external service or project source:

1. Confirm which account will be connected.
2. Check what information the connected service exposes.
3. Confirm the workspace allows the intended app or plugin.
4. Give Codex only the files and permissions required.
5. Review actions before approving changes to external systems.

OpenAI states that plugins and connected apps remain subject to provider authorization and workspace permissions, and that a plugin cannot access content beyond the permissions granted to the relevant account or managed source.

For shared Projects, remember that project members can access the shared project's chats and files. Shared projects use project-only memory and do not access a member's outside project memory.

Know When Not to Use Codex

Codex is powerful, but it is not always the best tool.

Use ChatGPT Work when the primary goal is:

• General web research
• Research reports
• Information synthesis
• Business analysis without technical execution
• Creating a document, spreadsheet, presentation, or report from research

Use Codex when the work benefits from:

• Code
• Repositories
• Local files
• Terminals
• Developer tools
• Browser-based execution
• Connected technical or business systems
• Long-running technical workflows
• Automation
• Building a tool or dashboard

This separation follows OpenAI's current positioning of Work and Codex.

Alternative Method: Use ChatGPT Work for Research, Then Codex for Execution

For many users, this is the cleanest workflow.

Phase 1 — Research in Work

Ask Work to investigate a topic, compare sources, and produce a structured research package.

Phase 2 — Store the useful context

Save the important findings and source material in the appropriate Project.

Phase 3 — Execute in Codex

Open Codex and use that context to:

• Analyze data
• Build tools
• Modify files
• Create technical prototypes
• Automate repetitive work
• Produce implementation changes

Phase 4 — Review

Check the final result before approving changes or external actions.

This approach uses each experience for the type of work OpenAI currently assigns to it instead of forcing one tool to perform every part of the process.

Common Problems and Fixes

Codex Is Missing

Problem: You cannot find Codex in the interface.

Fix: Make sure you are using the current ChatGPT desktop application. OpenAI currently provides Codex as a separate desktop experience; it is not selectable as a normal experience on ChatGPT web or mobile. Supported desktop Codex work can be accessed remotely from the mobile app, but the mobile app does not turn Codex into a normal mobile experience.

The Codex App Looks Different From an Older Tutorial

Problem: A guide online tells you to use an older Codex app layout.

Fix: Update the current desktop application. OpenAI has moved the previous Codex desktop experience into the new ChatGPT desktop application, which now combines Chat, Work, and Codex.

A Plugin Is Not Available

Problem: You cannot find the plugin used in a tutorial.

Fix: Plugin availability depends on plan, region, workspace settings, role, and the supported product surface. Check the Plugins directory and confirm that any required app is also available and authorized.

Codex Cannot Access a Business App

Problem: Codex can see the plugin but cannot retrieve or modify the required information.

Fix: Check the underlying app connection, provider authorization, workspace permissions, and the actions supported by that app. A plugin cannot bypass those restrictions.

Research Results Are Too General

Problem: The result sounds polished but does not answer the actual business question.

Fix: Rewrite the task around a concrete outcome, provide the required source boundaries, and define success criteria.

For example:

Do not give me a general overview.
I need a decision-ready analysis of whether this product should enter the SMB market.

Evaluate:

- Market size
- Competitors
- Pricing
- Customer needs
- Major risks
- Distribution options
- Evidence supporting each conclusion

Separate confirmed facts from assumptions.

Codex Tries to Act Too Early

Problem: The agent begins changing files or interacting with a connected system before you have reviewed the plan.

Fix: Explicitly separate analysis from execution:

First investigate and prepare a plan.
Do not modify files.
Do not send messages.
Do not change external systems.

Show me the proposed actions and wait for approval.

Common Mistakes to Avoid

Using Codex for Every Research Task

For straightforward research, Work may be the better experience. Codex becomes more valuable when research connects to files, tools, technical execution, automation, or a larger workflow.

Giving Codex a Vague Goal

"Analyze my business" is too broad.

Define what decision the analysis should support and what the finished result should contain.

Giving It Unrelated Context

Large volumes of irrelevant files can make a project harder to manage. Keep the context focused on the actual objective.

Trusting Unverified Claims

An AI-generated answer can still contain errors. Require evidence for important facts and review the final output.

Connecting Too Many Apps

Every connected service increases the amount of context and permissions involved. Connect only what the workflow needs.

Skipping Human Review

Do not automatically publish, deploy, send, purchase, or modify business data simply because Codex completed a task.

How to Check If It Worked

A successful Codex workflow should leave you with a result that matches the original objective.

Check these points:

1. The requested task was actually completed.
2. The output satisfies the success criteria you gave Codex.
3. Important claims have appropriate evidence.
4. Unsupported assumptions are clearly identified.
5. Files or project changes are where they should be.
6. No unexpected external actions occurred.
7. Any code or technical changes have been reviewed and tested where appropriate.
8. Business or connected-app actions were approved before execution.

For longer tasks, ask Codex to provide a final status report:

Give me a final verification report.
Include:
- What you completed
- What remains unfinished
- Files changed
- Sources used
- Tests or checks performed
- Known limitations
- Anything I should manually review

What to Do Next

Once you understand the basic workflow, start with one real task rather than a large automation project.

A good first Codex business workflow is something like:

Take these existing files and research notes.
Goal:
Turn them into a decision-ready competitive analysis.
Requirements:
- Keep factual claims traceable to sources.
- Identify missing information.
- Create a structured report.
- Create a comparison spreadsheet.
- Flag assumptions.
- Do not make external changes.
- Stop and ask for approval before executing anything outside this workspace.

That gives you a practical way to learn how Codex handles research, project context, connected tools, and review without giving it unnecessary authority.