Frequently Asked Questions

The FAQ below answers the recurring questions about Claude Code, Codex, and Cursor: which default model to start on, what the plans cost, how vendors handle your code and data, how to cut permission prompts safely, how to automate pull request review, and why usage limits run out. Each answer links to its in-depth page.

This page is for developers and tech leads who already have an agent installed and keep hitting the same small questions. A teammate asks whether they still have to read every diff, a flag from a blog post fails with “unexpected argument”, or the weekly limit runs out on a Wednesday. Every answer below was checked against claude --help (2.1.283 and later), codex --help (0.157.1), and the models hub. Answers that carry the date 2026-10-12 were re-checked that day; the others stand as of 2026-09-26.

With this page you get:

• A one-line model rule you can give the whole team instead of a model debate.
• The review and permission settings that let an agent work without you watching every step.
• Install and MCP commands for Claude Code and Codex, and the MCP config for Cursor.
• Three copy-paste prompts: prove a change without reading the diff, plan before editing, and find stale agent config in your repository.

How do I choose and pay for a tool?

What is the difference between Claude Code, Codex, and Cursor?

All three run the same kind of agent loop (read, edit, run commands, check), so the choice is mostly about where you work and which vendor’s models you want.

	Claude Code	Codex	Cursor
Where it runs	CLI first, plus desktop app, VS Code, JetBrains, web, and GitHub Actions	CLI, IDE extension, Codex desktop app, Codex cloud, and GitHub review	Editor first (a VS Code fork), plus Cloud Agents and a CLI
Default model	Claude Opus 5.5	GPT-6.1 Sol	Cursor names no single default model (checked 2026-10-12)
Project instructions	CLAUDE.md (falls back to AGENTS.md)	AGENTS.md	Rules; also reads AGENTS.md and CLAUDE.md
PR review	/code-review, managed Code Review, claude-code-action	@codex review, codex review	Bugbot

For the full decision, see how to choose between the tools.

Which AI model should I use?

Start on the tool’s default model, raise effort before you switch model, and switch model only when your own evals say so. On 2026-10-12 that means:

• Claude Code: Claude Opus 5.5 is the default on every paid plan from v2.1.280, on both release channels (checked 2026-10-12). It runs at medium effort; raise it with /effort high or /effort xhigh before you reach for another model.
• Codex: GPT-6.1 Sol is the bundled default since CLI 0.159.1 (2026-09-29). GPT-6 Luna is the cheap, fast option, and GPT-6 Astra is the pick for the hardest work.
• Longest-horizon work: pick Claude Fable 5.1 by hand (/model fable). It is never a default and may bill to usage credits.
• Cost-sensitive volume and subagents: Claude Sonnet 5.5, Claude Haiku 5.5 (in Claude Code from v2.1.293, the latest channel on 2026-10-12), or GPT-6 Luna.
• Cursor: Cursor names no single default model. Its Cursor Models pool is Grok 4.7, Grok 4.6, Grok 4.5 and Composer 2.5, and Cursor calls Grok 4.7 its flagship model (Cursor docs and help, checked 2026-10-12). Check the model picker in your own Cursor install.

Prices and context windows live on one page only: the models hub.

How much do the plans cost?

• Claude Code: the Free plan does not include it. Pro is $20 a month ($17 on annual billing), Max 5x $100, and Max 20x $200. Team Standard seats are $20 a month on annual billing ($25 monthly) and Team Premium seats $100 ($125 monthly). Enterprise is $20 per seat on annual billing plus usage at API rates. Limits run on a rolling 5-hour window and a weekly window.
• Codex: included in ChatGPT Plus, Pro, Business, Edu, and Enterprise plans. ChatGPT Pro has plans at $100, $200 and $500 a month (OpenAI’s pricing page, checked 2026-10-12). The $100 Pro tier gives 5x Plus usage and the $200 tier 20x (secondary: TechCrunch and 9to5Mac, 2026-04-09). Allowances also run on 5-hour and weekly windows.
• Cursor: Hobby is free; Pro is $20 a month, Pro+ $60 and Ultra $200; Teams is $40 per user per month for a Standard seat and $120 for a Premium seat (cursor.com/pricing, checked 2026-10-12). Start is a separate plan sold in India only.

For team budgets and scenarios, see pricing analysis and cost governance.

Can I use Claude Code, Codex, and Cursor in the same repository?

Yes. Each tool reads its own instruction file, so the risk is drift between files, not conflict between tools:

1. Keep one source of truth: AGENTS.md. Codex reads it before it starts work.
2. Claude Code reads CLAUDE.md, and reads AGENTS.md when a project has no CLAUDE.md (v2.1.277 and later, both release channels on 2026-10-12). Keep CLAUDE.md for Claude-only notes, or delete it.
3. Cursor uses its own Rules. Put a short rule there that points at your single source-of-truth file.
4. Add a CI check that fails when the build, test, or lint commands in those files disagree with package.json or your Makefile.

How do vendors handle my code and data?

Training use depends on plan and contract, and this page does not restate it; get it in writing (four questions below). Verified retention and policy details as of 2026-09-26, except where a bullet names 2026-10-12:

• Claude (retention): Claude Fable 5.1 is a Covered Model and retains data by default. Organizations with Zero Data Retention cannot use managed Code Review or Ultrareview.
• Codex (policy control): administrators can pin policy in requirements.toml, including enforce_residency.
• Cursor: check the Privacy Mode setting and whether your admin enforces it. For teams, Privacy Mode is enabled by default for all team members, and admins can enforce it org-wide (Cursor docs, checked 2026-10-12).

Before you roll out, ask each vendor four questions in writing: is our code used for training, how long is it retained, which models need retention, and can we enforce the setting for every seat. The privacy and data handling guide and the procurement questionnaire turn those into a policy.

How do I trust what the agent ships?

Do I still review every line of agent-written code?

No. Line-by-line review cannot keep up once agents write most of the diff, so move your attention to evidence:

1. Feedback loop: the agent runs tests, the type checker, and the linter before it says “done”.
2. Acceptance criteria: every change maps to a criterion you wrote before the work started.
3. Review agent: a second model reviews the pull request (see the next question).
4. Human sign-off: you own intent, risk, the merge, and production.

Read every line only for regulated or critical code, or when a check or review finding says the change can break behavior, leak data, or breach policy. The AI-native lifecycle and the human in the loop explain where each check sits.

Copy-paste prompt: prove a change without me reading the diff

Before you tell me this change is done, prove it. Produce a short evidence report:
1. List each acceptance criterion from the task and the test that covers it. Name the file and test.
2. Run the full test suite, the type checker, and the linter. Paste the exact commands and the final summary lines.
3. List every file you changed that touches authentication, payments, data deletion, or database migrations. For each one, say what could break and which test would catch it.
4. List anything you could not verify, and why.
Do not summarize the diff. If a check fails, fix it and rerun before you report.

How do I get pull requests reviewed automatically?

Claude Code

• In a session, run /code-review for a local review, or /code-review ultra for a cloud-hosted multi-agent review. From the shell, claude ultrareview reviews the current branch or a PR number.
• On Team and Enterprise, the managed Code Review (generally available per Anthropic, checked 2026-10-16) reviews pull requests and reads a repository-root REVIEW.md for review-only rules. Local /code-review does not read REVIEW.md.
• In CI, use anthropics/claude-code-action@v1.

Codex

• Comment @codex review or @codex security review on a GitHub pull request, and write review rules in AGENTS.md.
• Locally, run /review in a session, or codex review --base main (also --uncommitted and --commit <SHA>) from the shell.

Cursor

• Bugbot reviews pull requests and flags bugs, security issues, and code quality problems. Set it up by following Cursor’s Bugbot documentation. Bugbot bills by usage on Teams and Individual plans, not as a per-seat subscription (Cursor, checked 2026-10-12).

A review agent is one layer, not the gate. Keep tests and CI required, and let a human decide the merge. See reviewing agent pull requests for the layered setup.

How do I stop permission prompts without giving the agent full access?

Allowlist what you trust; do not switch checks off.

Claude Code

From v2.1.283 (both release channels on 2026-10-12), auto mode is the starting mode for interactive terminal and VS Code sessions: a classifier reviews each action instead of prompting you. claude -p and the Agent SDK still start in Manual, except on third-party providers or with telemetry off, where they start in auto mode from v2.1.285. Sessions on models auto mode does not support (such as Haiku 4.5) also start in Manual.

To pre-approve specific commands, add them to .claude/settings.json:

{
  "permissions": {
    "allow": ["Bash(npm test)", "Bash(npm run lint)"]
  }
}

Use --dangerously-skip-permissions only inside a disposable, network-restricted container.

Codex

Approval policy and sandbox are separate controls. on-request is the default: Codex asks only when it needs to go past the sandbox. --approve-for-me routes those requests to an automatic reviewer with the workspace-write sandbox. OpenAI’s preferred sandbox control is a permission profile (beta), for example -c default_permissions=":workspace".

-a never means Codex never asks, so pair it only with a sandbox you chose on purpose. Use --dangerously-bypass-approvals-and-sandbox only in an environment that is already sandboxed.

Cursor

Cursor groups these controls under Run modes in its agent security settings. You find them under Settings > Agents > Approvals & Execution: Auto-review (the recommended default since Cursor 3.6), Allowlist and Run Everything (Cursor docs, checked 2026-10-12). Cloud Agents do not use Run modes, because they run in their own dedicated machine.

For the full model, see permissions and sandboxing.

How do I make the agent think harder on a difficult task?

Raise effort, not the model. In Claude Code, run /effort high, /effort xhigh, or /effort max in the session, or start with claude --effort xhigh. /effort ultracode turns on automatic workflow orchestration at the session’s current effort level (v2.1.284+); claude --effort ultracode also starts at xhigh. In Codex, GPT-6.1 Sol runs at low effort by default; raise it as far as max, or to ultra, a Codex-only level that also delegates tasks automatically (GPT-6 Luna has no ultra). In Claude Code, typing the keyword ultracode in a prompt has the same effect as /effort ultracode; prefer the command so the level is visible. Effort defaults per model are on the models hub.

How do I set up the tools?

How do I install Claude Code and Codex?

Claude Code

# Native installer: auto-updates, tracks the latest channel
curl -fsSL https://claude.ai/install.sh | bash
claude --version

The Homebrew cask claude-code tracks the stable channel, which runs about a week behind. To follow the latest channel with Homebrew, install claude-code@latest instead. See installing Claude Code.

Codex

npm install -g @openai/codex
# or, on macOS
brew install --cask codex
codex --version

See installing Codex.

Cursor

Download the editor from cursor.com and sign in. See installing Cursor.

How do I add an MCP server?

MCP (Model Context Protocol) servers give the agent tools: a browser, a database, your issue tracker. This example adds the Playwright browser server (@playwright/mcp, version 0.0.83 on npm on 2026-09-26).

Claude Code

claude mcp add playwright -- npx -y @playwright/mcp@latest
claude mcp list

Add -s project to write the server to .mcp.json so the team shares it.

Codex

codex mcp add playwright -- npx -y @playwright/mcp@latest
codex mcp list

This writes an [mcp_servers.playwright] table to ~/.codex/config.toml.

Cursor

Add the server to .cursor/mcp.json in the project:

{
  "mcpServers": {
    "playwright": { "command": "npx", "args": ["-y", "@playwright/mcp@latest"] }
  }
}

Start with two or three servers; each one adds tool definitions to the agent’s context. See the MCP server directory and essential servers.

Should I use an MCP server or an Agent Skill?

Use a skill when the agent needs know-how: a procedure, a checklist, or a house style in a SKILL.md folder. The agent reads only the name and description until the skill fires, so skills cost little context. Install one with npx skills add <owner/repo> (the skills CLI from vercel-labs/skills), which targets Claude Code, Codex, Cursor, and other agents.

Use an MCP server when the agent needs live access to a system: a database, a browser, an API. It runs as a process or a remote endpoint and exposes tools on every turn. If both exist for a workflow, start with the skill and add the server when the skill has to reach real data. See installing and managing skills.

How do I work day to day?

Should the agent plan before it edits?

Yes, for anything that touches more than one file. All three tools have a plan mode: /plan in Claude Code, /plan in Codex, and Plan Mode in Cursor. Review the plan against your acceptance criteria, then let the agent execute. The artifact chain describes the full sequence from intent to plan to tasks, and agent vs ask mode covers when a read-only session is the better choice.

Copy-paste prompt: plan before editing

Do not edit any files yet. Read the code involved in this task and write a plan:
1. Restate the goal in one sentence and list the acceptance criteria you will be checked against.
2. List every file you expect to change and why.
3. List the tests you will add or change, and the command that runs them.
4. Name the riskiest step and how you will know it worked.
5. List open questions for me. Stop after the plan and wait for my approval.

Why did I hit my usage limit so fast?

Usage grows with the model, the effort level, the size of the context, and the number of tool calls. A long session that re-reads big files costs far more than several short ones. To stretch the same plan:

1. Check where usage goes with /usage (Claude Code and Codex).
2. Start a fresh session between unrelated tasks: /clear in Claude Code, /new in Codex.
3. Summarize a long session with /compact (both tools).
4. Send volume work and subagents to a cheaper model before you lower effort on the main task.

For the mechanics, see the cost of context.

What breaks when you follow an older FAQ?

Agent tools change weekly, so advice from a few months ago often fails in ways that look like your mistake.

• codex --full-auto or codex exec --full-auto fails. The flag is gone from codex --help and codex exec --help in 0.157.1. Recovery: use a permission profile (-c default_permissions=":workspace"), --approve-for-me, or --sandbox workspace-write.
• codex -a on-failure fails with “invalid value”. In 0.157.1, -a accepts only on-request and never. codex exec -a never fails too, because exec has no -a; write codex -a never exec "…".
• Opus 5.5 is missing on an old install. Opus 5.5 needs v2.1.280 or later, and the claude-code Homebrew cask tracks the stable channel, which is v2.1.292 on 2026-10-17. Recovery: check claude --version; on an older build, upgrade the cask, install claude-code@latest or use the native installer.
• The GitHub MCP server will not start. @modelcontextprotocol/server-github is deprecated on npm. Recovery: use GitHub’s hosted server; see the GitHub MCP guide.
• Cursor setup guides still point at .cursorrules. Configure Cursor with Rules instead of a root .cursorrules file.

Copy-paste prompt: find stale agent config in this repository

Audit this repository for outdated AI agent configuration. Search CLAUDE.md, AGENTS.md, .cursor/, .claude/, .codex/, .mcp.json, CI workflows, Makefiles, and package.json scripts for:
- codex --full-auto, codex -a on-failure, codex -a untrusted, or codex exec -a
- @modelcontextprotocol/server-github or any MCP server using the SSE transport
- a .cursorrules file (move its content to Cursor Rules)
- hardcoded model names older than the current defaults
For each hit, show the file and line, explain what fails today, and propose the replacement. Do not change anything until I confirm.

Where to go next from this FAQ

• New to agents: follow the developer track from the first session to running agents unattended.
• Choosing a tool: read the comparison hub and the models hub.
• Something is failing: the troubleshooting guide covers error messages; the glossary defines the terms used here.
• Still stuck: ask in the community.