Tool map: Claude Code, Cursor, and Codex

The tool map keeps intent.md, spec.md, plan.md, verification evidence, review findings, and incident records portable across Claude Code, Cursor, and Codex. The lifecycle and human gates remain identical; only the harness surface changes. Use the table to choose an implementation without duplicating or weakening the canonical stage procedure.

Commands and product surfaces were checked on 4 September 2026 against the installed CLIs and official vendor documentation. Model names and prices are deliberately absent: route by evaluated task quality, latency, and completed-work cost, then check the current picker or vendor pricing page.

Current capability map

Play	Claude Code	Cursor	Codex
Capture intent	Interactive repository session or Claude web surface; commit the reviewed intent.md	Ask mode for discovery; Agent mode to save the accepted file	Interactive /plan or app chat; commit the reviewed intent.md
Spec from intent	Project skills in .claude/skills/ plus repository context in CLAUDE.md	Project skills in .cursor/skills/ or .agents/skills/; rules in .cursor/rules/*.mdc	Project skills in .agents/skills/; repository guidance in AGENTS.md
UI evidence	Image input, browser or design MCP chosen by the team	Image input, browser tools, or Figma MCP chosen by the team	Image input, browser/computer use, or an approved MCP tool
Plan before build	--permission-mode plan, /plan, or Shift+Tab; save the accepted plan in the repo	Plan mode in IDE or cursor-agent --plan; save the accepted plan in the repo	/plan before the active run; save the accepted plan in the repo
Parallel isolation	claude -w NAME or explicit Git worktrees	Agents Window worktrees, /worktree, /best-of-n, or explicit Git worktrees	Worktree in the desktop app or explicit Git worktrees; no CLI --worktree flag in 0.146.0
Deterministic controls	Hooks in .claude/settings.json with scripts under .claude/hooks/; permission modes	Project or enterprise hooks, sandboxing, rules, and worktree setup	Hooks in .codex/hooks.json or config.toml; sandbox and approval policy
Session feedback	Commands in CLAUDE.md; tests, builds, browser and screenshots	Commands in rules; terminal, browser, checkpoints, screenshots	Commands in AGENTS.md; tests, terminal, browser and screenshots
Non-interactive run	claude -p with explicit permission and output format	cursor-agent --print with sandbox and output format	codex exec with explicit sandbox; --json or --output-schema when needed
Pull-request review	Managed Code Review or claude-code-action	Bugbot on the PR; /review and /review-bugbot on supported surfaces	codex review locally or openai/codex-action@v1 in CI
Background work	Background sessions, routines, CI schedules, or the Agent SDK	Cloud Agents and Automations	App scheduled tasks, cloud tasks, CI schedules, Slack or Linear intake
Production authority	External release gate and credentials unavailable to the agent	External release gate and credentials unavailable to the agent	External release gate and credentials unavailable to the agent

Similar names do not imply identical controls

A rule, skill, hook, sandbox, and approval gate solve different problems. Rules and skills steer behavior. Hooks and platform policies can enforce deterministic decisions. A Git worktree isolates files but not production databases, cloud accounts, secrets, queues, or third-party APIs.

Verified command surface

# Claude Code 2.1.258
claude --permission-mode plan
claude -w feature-slug
claude -p --output-format json "Review the current diff."

# Cursor Agent 2026.01.28-fd13201
cursor-agent --plan "Propose an implementation plan."
cursor-agent --print --sandbox enabled --output-format json "Review the current diff."

# Codex CLI 0.146.0
codex
# Then enter /plan before the agent starts working.
codex review --uncommitted
codex exec --sandbox workspace-write --json "Run the documented checks."

Verify against the local installation before automation:

claude --version && claude --help
cursor-agent --version && cursor-agent --help
codex --version && codex --help
codex exec --help
codex review --help

Choose the implementation

Claude Code adapter Plan mode, CLAUDE.md, skills, hooks, worktrees, subagents, and print mode.

Cursor adapter Ask, Plan, and Agent modes, rules, worktrees, Bugbot, Cloud Agents, and Automations.

Codex adapter Plan mode, AGENTS.md, skills, app worktrees, sandboxing, review, and scheduled tasks.

When mapping fails

A vendor has no equivalent control. Keep the control in Git, CI, identity, or the deployment platform. Do not rename an advisory prompt and present it as enforcement.

A copied CLI flag fails. Compare the command with the installed --help output and the dated verification note above. Update both language versions and the relevant adapter.

A tool-specific artifact becomes the only source of truth. Copy the accepted result into the repository or explicitly link the authoritative external record and commit SHA.

Continue with the lifecycle

Start with the artifact chain, adopt the sequence through Plan, and measure it with SDLC metrics.