Where Cursor, Claude Code and Codex actually differ.
Most rows read yes in all three. The ones that differ are where the agent runs, which controls block an action, and what checks a result before merge. AI Developer Toolkit teaches all three side by side: 400+ guides and 100+ recipes, each showing when it was last updated.
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
- Cursor docs
- 2026-08-28
- Claude Code 2.1.283
- 2026-09-26
- Codex 0.157.1
- 2026-09-26
- List prices captured
- 2026-09-26
Six stations, three tools.
Each station is a job a coding-agent workflow has to do. Each cell lists what the tool’s own documentation or CLI names for it.
Intent
Project instructions, skills, plan mode
Same shape All three read a project instruction file, install skills and plan before code; the file name changes: CLAUDE.md in Claude Code, AGENTS.md in Codex, Rules and AGENTS.md in Cursor.
Cursor
- Rules
- Cursor's docs:
AGENTS.mdin the project root and in subdirectories (nested files combine, the more specific one wins), andCLAUDE.mdread the same way and always applied;.cursorrulesis legacy Cursor token management guide, 2026-10-12 - The agent runtime in
@cursor/sdk1.0.32 loads a root and nestedAGENTS.mdunconditionally, andCLAUDE.mdonly with third-party config loading on, while Cursor's docs sayCLAUDE.mdis always applied Cursor token management guide, Cursor SDK 1.0.32, 2026-09-26 - Agent Skills
- Plan Mode
Cursor docs, 2026-08-28
Claude Code
-
CLAUDE.md; by default readsAGENTS.mdonly when there is noCLAUDE.md(v2.1.277+), and the/configProject instructions setting loads both - Skills in
.claude/skills/; bundled/code-review,/batch,/simplify - Plan mode:
/plan, Shift+Tab,--permission-mode plan
Claude Code 2.1.283, 2026-09-26
Codex
-
AGENTS.md(/initcreates one) - Skills:
/skills,.agents/skills/ - Plan:
/plan
Codex 0.157.1, 2026-09-26
Harness
Permissions and approvals, sandbox, MCP, hooks, organisation policy
Differs Enforcement depth differs: 33 documented hook events in Claude Code against 12 in Codex, and Read deny rules in Claude Code's settings; Cursor's hooks can block, its agent runtime refuses reads of .cursorignore paths, and its run modes are Auto-review, Allowlist and Run Everything (Cursor docs, checked 2026-10-12).
Cursor
- Run modes: Auto-review (the recommended default), Allowlist and Run Everything Cursor docs, Cursor 3.24, 2026-10-12
- Auto-review runs shell commands in a sandbox when possible;
sandbox.jsonconfigures it Cursor docs, Cursor 3.24, 2026-10-12 - Hooks: JSON over stdio, before or after agent-loop stages; can block
- MCP
- Organisation policy: Team and MDM settings layers (SDK 1.0.32)
- File deny rules: the agent runtime refuses reads of paths in
.cursorignore, passes the list to the shell sandbox and blocks the agent from editing the file Cursor privacy and security guide, Cursor SDK 1.0.32, 2026-09-22 - Starting mode: Auto-review, the recommended default since 3.6 (2026-05-29) Cursor docs, Cursor 3.24, 2026-10-12
Cursor docs, 2026-08-28
Claude Code
- Permission modes: Manual (
default),acceptEdits,plan,auto,dontAsk,bypassPermissions; interactive sessions start in auto mode from v2.1.283,claude -pin Manual (auto mode there only on third-party providers or with telemetry off, from v2.1.285) - Sandboxed Bash: macOS Seatbelt, Linux and WSL2 bubblewrap
- Hooks: 33 events;
command,http,mcp_tool,promptandagenthandlers - MCP client:
claude mcp add(--scope projectwrites.mcp.json); runs as an MCP server withclaude mcp serve -
Readdeny rules in.claude/settings.json, for exampleRead(./.env) - Organisation policy: managed settings (
availableModels,permissions.disableAutoMode,deniedModels; v2.1.283)
Claude Code 2.1.283, 2026-09-26
Codex
- Approval policy (
-a on-requestor-a never) plus automatic review (--approve-for-me); starts inon-request - OS sandbox:
--sandbox read-only,workspace-write,danger-full-access; permission profiles (beta) - Hooks: 12 events; need persisted trust before they run
- MCP client:
codex mcp add,config.toml; does not run as an MCP server (removed in 0.154.0) - File access: a permission profile with a
denyfilesystem rule (beta);--sandboxalone restricts writes, not reads - Organisation policy:
requirements.toml(allow_managed_hooks_only,permission_profile,mcp_servers)
Codex 0.157.1, 2026-09-26
Loop
Stop conditions, /goal, /loop, scheduled runs, headless
Differs All three run headless and on a schedule, but Claude Code Routines and Cursor Automations run in the cloud while Codex automations run locally in the desktop app. /goal and /loop differ (Codex has no /loop in 0.162.1; Cursor's /goal is marked "Rolling out"), and Claude Code can also serve as an MCP server, which Codex removed in 0.154.0.
Cursor
-
/goal, marked "Rolling out" Cursor docs, Cursor 3.24, 2026-10-12 -
/loopbundled skill since 3.5 - Scheduled runs: Automations, which start Cloud Agents
- Headless: print mode (
agent -p) Cursor docs, Cursor 3.24, 2026-10-12 - Machine-readable output:
--output-format textis documented Cursor docs, Cursor 3.24, 2026-10-12 - Other output formats: not verified for this page (checked 2026-10-12) Cursor docs, Cursor 3.24, 2026-10-12
Cursor docs, 2026-08-28
Claude Code
-
/goal, judged by a separate evaluator model each turn -
/loopbundled skill (alias/proactive) - Scheduled runs: Routines in the cloud (
/schedule, research preview) - Headless:
claude -p;--output-format jsonorstream-json;--json-schema
Claude Code 2.1.283, 2026-09-26
Codex
-
/goal(featuregoals, stable) - No
/loopin Codex 0.162.1 (slash-command enum, binary strings) or the developer-commands reference (checked 2026-10-12); use/goalor CI Loop command guide, Codex 0.162.1, 2026-10-12 - Scheduled runs: automations run locally in the desktop app, which must be open and the machine awake Background and cloud agents guide, 2026-09-26
- Headless:
codex exec(aliase);--json(JSONL events);--output-schema FILE;-o FILE
Codex 0.157.1, 2026-09-26
Graph
Subagents, worktrees, background sessions
Differs Where the agent runs differs: in the editor (Cursor), from the terminal out to IDE, desktop and web (Claude Code), or spread across CLI, IDE, desktop app, cloud tasks and Slack and Linear hand-offs (Codex).
Cursor
- Subagents, including on their own cloud VMs
- Worktrees
- Background sessions: Cloud Agents in isolated VMs
Cursor docs, 2026-08-28
Claude Code
- Subagents: built-in Explore, Plan and general-purpose; custom ones in
.claude/agents/ - Worktrees:
--worktree [name](-w) - Background sessions:
--bg, thenclaude attach,logs,stop - Cloud: web sessions (
claude --cloud)
Claude Code 2.1.283, 2026-09-26
Codex
- Subagents:
/subagents; multi-agent on by default - Worktrees:
--worktreeflag or/worktree - Background sessions: the
codex agentscommand center - Cloud: Codex cloud (
codex cloud, experimental)
Codex 0.157.1, 2026-09-26
Verification
Diff review, review agents, security review, checkpoints and rewind
Same shape All three review in the session and rewind. Claude Code has /code-review, /security-review and checkpoints; Codex has /review; Cursor has Agent Review, /review and /review-security. One caveat: Codex's Esc Esc backtracks the conversation but does not restore files (0.157.1), so file state is Git's job.
Cursor
- In the session: Agent Review (
/agent-review) and/review, which runs/review-bugbotCursor code review guide, 2026-09-15 - Bugbot on the pull request
- Security review:
/review-securitybefore the push; Security Agents on the pull request Cursor code review guide, 2026-09-15 - Rewind: checkpoints, stored locally and separate from Git; restoring one reverts files only Cursor docs, Cursor 3.24, 2026-10-12
Cursor docs, 2026-08-28
Claude Code
- In the session:
/code-review(--fix,--comment) - From the shell or CI:
claude ultrareview(cloud, multi-agent); managed Code Review (Team and Enterprise; generally available per Anthropic) Code review automation guide, 2026-10-16 - Security review:
/security-review - Rewind: checkpoints and
/rewind(Esc Esc)
Claude Code 2.1.283, 2026-09-26
Codex
- In the session:
/review - From the shell or CI:
codex review,codex exec review - Security review:
@codex security review - Rewind: Esc Esc backtracks the conversation to an earlier message; it does not restore files (0.157.1). Use Git for file state
Codex 0.157.1, 2026-09-26
Release
CI and GitHub Actions, repo-event triggers, pull-request routing
Differs Review on the pull request differs: Codex reviews on a mention (@codex review), Cursor adds Bugbot, Security Agents and reviewer routing, Claude Code reviews through its GitHub Action or managed Code Review.
Cursor
- GitHub Actions: Cursor CLI in GitHub Actions (documented)
- Review on the pull request: Bugbot
- Hand-offs: Automations triggered from Slack, Linear, GitHub, GitLab, webhooks, Sentry and PagerDuty
- Reviewer routing and approval: PR Routing & Approval (can approve low-risk PRs)
- Start a cloud run from code: Cloud Agents API and webhooks
Cursor docs, 2026-08-28
Claude Code
- GitHub Actions and review on the pull request:
anthropics/claude-code-action@v1 - Hand-offs: Slack (Claude Tag for Team and Enterprise); others through MCP
- Reviewer routing and approval: through your own CI and CODEOWNERS
- Start a cloud run from code: Routines with an API trigger (research preview)
Claude Code 2.1.283, 2026-09-26
Codex
- GitHub Actions:
openai/codex-action@v1 - Review on the pull request:
@codex reviewon GitHub and GitLab, on a mention and without a separate workflow file - Hand-offs: Slack (
@ChatGPT, with Cloud delegation to a Codex Cloud task) and Linear (@Codex) Codex docs, 2026-10-12 - Reviewer routing and approval: through your own CI and CODEOWNERS
- Start a cloud run from code:
codex cloud(experimental)
Codex 0.157.1, 2026-09-26
Read a gap as a gap in what we read. A cell that says “not found in the docs” means we could not source it on that date. It does not say the tool lacks the feature.
- Drawing
- Six stations, three tools
- Source
- The feature matrix guide, last updated 2026‑10‑17
- Checked
- Cursor docs, 2026-08-28. Claude Code 2.1.283, 2026-09-26. Codex 0.157.1, 2026-09-26.
Start from the situation you are in.
Five situations, and the tool the feature matrix points to for each. Before you buy for a team, test the rows that decide it on your own repository.
-
WhenThe team works in a VS Code-style editor and wants to steer each change
Start with Cursor Tab, inline edit, Plan Mode and the agent in one window
-
WhenThe work should run with less supervision, gated by hooks and CI
Start with Claude Code 33 hook events, deny rules,
claude -pwith a schema -
WhenOne agent across desktop, CLI, IDE, cloud tasks, Slack and Linear
Start with Codex Multi-surface by design; review on
@codex review -
WhenPull-request review and reviewer routing matter most
Start with Cursor or Codex Bugbot and PR routing, or review on a mention
-
WhenAn editor agent for interactive work plus a CLI agent in CI
Start with Two tools Both keep instructions in files
The five comparisons.
Each one opens with a plain verdict, then goes station by station, and quotes prices with the date they were captured.
How this page stays honest.
Claude Code ships roughly one release a day, so any comparison ages fast. This one tells you how old each cell is, and what we did not find.
# what a model suggested codex -a on-failure # what Codex 0.160.1 prints error: invalid value 'on-failure' for '--ask-for-approval <APPROVAL_POLICY>' # what Codex accepts codex -a on-request
-
Every cell shows when it was checked.
Claude Code and Codex cells come from the installed CLIs, Claude Code 2.1.283 and Codex 0.157.1, on 2026‑09‑26. Cursor cells come from its documentation, read on 2026‑08‑28. An item the docs read again later carries its own day.
-
A gap is a gap in what we read.
When we cannot source a cell it says “not found in the docs”, with the date we looked. It never says “not supported”.
-
Prices carry a snapshot date.
Public US list prices, captured 2026‑09‑26; a row read later says so in its note. Regional prices, negotiated contracts and promotions can differ, so check the vendor page before you buy.
-
Your own install outranks this table.
The feature matrix guide says to run the version check on your machine and trust that
--helpoutput over any table, including this one.
Whichever you pick, set it up once.
Using several tools means explaining the same project again. The Setup Pack asks nine questions about your stack and generates the instruction files for all three in about 20 minutes. Your own agent then runs INSTALL-PROMPT.md in your repository to verify every command and server before it writes a line.
“Using several of them can feel messy. I switch between Claude Code, Cursor, VS Code, GitHub, and other agents, and often have to explain the same project or problem again.” What the pack generates
One AGENTS.md for every agent, a CLAUDE.md that points at it, and Cursor rules. It generates a zip; it does not write into your repository.
- AGENTS.md
- CLAUDE.md
- .claude/rules/testing.md
- .cursor/rules/testing.mdc
- .mcp.json
- .cursor/mcp.json
- TESTING.md
- INSTALL-PROMPT.md
- SETUP.md
- README.md
Included in the Toolkit, with a free preview for one tool.
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
Questions before you choose.
Something else? Write to support@developertoolkit.ai.
Which of the three has the most features?
None of them wins every row. On 2026-10-12 all three have subagents, worktrees, skills, hooks, MCP, a headless mode, an SDK and scheduled runs (Routines in the cloud for Claude Code, Automations in the cloud for Cursor, scheduled tasks and local automations in the desktop app for Codex). They differ in where the agent runs, how guardrails are enforced and what review runs on a pull request, so test the rows that decide your choice on your own repository.
Which one should I start with?
Claude Code comes with Claude plans and Codex with ChatGPT plans, so a subscription you already pay for may decide it. Otherwise follow the matrix by situation: Cursor if the team works in a VS Code-style editor and wants to steer each change; Claude Code if the work should run with less supervision, gated by hooks and CI; Codex for one agent across desktop, CLI, IDE, cloud tasks, Slack and Linear.
How much do they cost?
On 2026-10-12 Claude Code is included in Claude Pro ($20 a month), Max 5x ($100) and Max 20x ($200), and in Team seats from $20 a seat billed annually. Codex is included in ChatGPT Plus ($20 a month), Pro ($100, $200 or $500) and Business ($20 per user per month billed annually), and in Edu and Enterprise (OpenAI’s pricing page, read 2026-10-12); the 5x and 20x Plus-usage ratios of Pro at $100 and $200 come from a secondary source, TechCrunch, 2026-04-09. Cursor Pro is $20 a month, Pro+ $60 and Ultra $200, and Teams Standard and Premium are $40 and $120 per user per month (cursor.com/pricing, read 2026-10-12); Cursor’s included-usage amounts per plan are not verified here. Every plan meters agent work beyond the seat, so the seat price is a floor.
Do I have to pick just one?
No. The last situation in the feature matrix is an editor agent for interactive work plus a CLI agent in CI, which is two tools, and both keep their instructions in files. Each tool is sold on its own plan, so ask what the second one has that the first lacks before you add a seat.
How current is this page?
Each cell carries its own date. Claude Code cells were checked against Claude Code 2.1.283 and Codex cells against Codex 0.157.1 on 2026-09-26; Cursor cells with no later stamp come from its documentation as read on 2026-08-28, because cursor.com could not be reached on 2026-09-26. Items stamped 2026-10-12 were re-read that day: for Cursor (Cursor 3.24) the run modes, the starting mode, the sandbox, checkpoints, /goal, AGENTS.md reading and print mode (agent -p); for Codex the Slack and Linear hand-off and the /loop check; for Copilot the status of local sandboxing. Claude Code ships roughly one release a day, so re-check a row before you automate against it.
Can’t I just ask an AI to compare them?
You can, and the answer will sound right. A flag that reads correctly may be one your release rejects: on Codex 0.160.1, codex -a on-failure fails with “error: invalid value 'on-failure' for '--ask-for-approval <APPROVAL_POLICY>'”, and codex -a on-request is what the flag accepts (Codex version history, checked 2026-10-10). Run --help on your own install and trust that over any table, this one included.
Does this page answer your questions?
One tap, no sign-in.
What are you here for?
What is holding you back?
How does it read?
Thanks. If you left an email, Michał will reply.
Answering your questions?
One tap each. Close whenever you like.
no sign-in
Pick your tool. Then climb a rung in it.
The guides cover Claude Code, Codex and Cursor side by side, in the same order, so a move to another tool later costs you little.
$59.99 once for 12 months of access. Nothing renews on its own. 30-day refund on a first purchase.
Michał replies, not a bot. The address is used for this reply only.