AI tools for WordPress developers in 2026 — what actually works

Most AI tool lists are useless because they are not written for anyone in particular. A list for "developers" that includes a design tool and a writing assistant tells you nothing. This covers what works specifically for WordPress plugin development and client project work — with actual limitations, not marketing copy.

Claude Code for PHP and JavaScript refactoring

Claude Code is the strongest tool available for multi-file refactoring tasks. The context window comfortably loads an entire plugin directory and reasons about changes that span many files at once. For WordPress-specific work:

Extracting classes from bloated plugin files. Plugin codebases accumulate logic in the main file over time. Pointing Claude Code at the file and describing the target structure produces a clean extraction — class definitions in their own files, require_once updates, namespacing, and accurate phpdoc blocks. The output is usually ready to commit with light review.

Debugging block editor component issues. Stale closures, missing dependency arrays, setAttributes calls that do not update because of a referential equality issue — Claude spots these reliably. It also handles harder cases now: race conditions between useEffect and useSelect, blocks that fail validation after a refactor, deprecation paths between block API versions.

Building blocks from scratch. Describe a block — attributes, allowed inner blocks, sidebar controls, save behaviour — and Claude generates a working block.json, edit.tsx, save.tsx, and view.js with the patterns it actually should follow (dynamic vs static, useInnerBlocksProps, asset splitting). The block-API-specific knowledge has caught up.

Where it still needs review. Two areas: (1) security defaults — capability checks on REST routes, nonce verification on AJAX, escaping on output. Claude knows these patterns but sometimes omits them in a quick draft. Always grep for current_user_can, wp_verify_nonce, and esc_* before merging. (2) Plugin-specific conventions — if your codebase uses a custom service container, a specific hook prefix, or a non-standard build pipeline, Claude defaults to generic patterns until you tell it otherwise. A short CLAUDE.md in the repo solves most of this.

The mental model is no longer "fast first draft." It is a competent collaborator that needs the same code review you would give a mid-level developer — faster on the structural work, faster on the boilerplate, and still wrong sometimes on security and project-specific conventions.

Claude Code vs Cursor vs Copilot — what each is for

The tool landscape settled in 2025 around three modes of use:

GitHub Copilot — line-level autocomplete inside the editor. Best when you are actively typing and want predictions one line ahead. A WP_Query loop, a settings field registration, a register_rest_route skeleton — Copilot completes these accurately and stays out of the way.

Cursor — chat-style assistant tightly integrated with file context in your editor. Best for small-to-medium edits where you want to keep your hands on the cursor but get help across two or three files.

Claude Code — agentic terminal-based assistant that reads, plans, edits, and tests across the whole codebase. Best for tasks that need multi-file reasoning, running test suites, or driving a longer refactor end-to-end.

For WordPress work specifically, Claude Code wins on plugin-wide refactors and Cursor wins on single-block edits. Copilot complements both — it covers the moments you are writing fresh code line by line.

Hook discovery used to be a weak spot for all of these tools. In 2026 it is not — modern models handle the core hook firing order (init, wp_loaded, template_redirect, pre_get_posts, etc.) and priority interactions accurately. Where they still need help is third-party plugin hooks and very recent core additions; for those, point them at the source or the WordPress Developer Resources.

AI for client proposals and communication

This is underused by most developers. Writing a project proposal from scratch takes an hour. Describing the project scope, stack, timeline, and constraints in a few sentences and asking for a draft takes five minutes. The structure and tone are usually 70% usable, and editing a draft is faster than writing from blank.

The same applies to scope change conversations. "The client added requirements that increase the project by 40% and I need to address this in writing" is an uncomfortable email. Describing the situation and asking for a professional draft gives a starting point. The key is specificity — vague prompts produce generic output. Describe the client's communication style, the exact change, and what outcome the email is trying to achieve.

AI for ACF field structure planning and dummy content

Two concrete WordPress time-savers:

ACF field group planning. Describe what a client wants to be editable in the backend — "a team member post type with name, title, bio, headshot, and LinkedIn URL" — and ask for the ACF field group structure as a JSON export. The output is not always immediately importable, but it maps the field types and structure correctly for straightforward cases.

Test data generation. Building out a new custom post type requires realistic example content to test pagination, layout, and edge cases. Asking for 20 realistic team members or testimonials as an array of objects — suitable for WP-CLI import or a seeding script — produces usable content in seconds. For generic content types this works well; for domain-specific content it needs editing.

Where AI still needs a human in the loop

Most of the "AI cannot do X" claims from a year ago no longer hold. The honest 2026 list is shorter and more specific.

Site-specific performance diagnosis. AI explains why posts_per_page: -1 is dangerous (no LIMIT, full result set in memory, no pagination) and recommends architectural fixes — WP_Query with proper pagination, tax_query JOIN behaviour, the meta_query indexing trap. What it cannot do is run EXPLAIN on your database and read the actual query plan. For "this specific page is slow on this specific site," you still need Query Monitor, New Relic, or the slow query log. AI takes over once you have the data.

Plugin compatibility debugging without the code. Give Claude Code access to two conflicting plugins and it will reliably spot filter priority collisions, global variable clashes, and function name overlaps. The constraint is access — most installed-plugin conflicts involve commercial plugins whose source is not in your repo. Once you provide the relevant files, AI is genuinely good at this.

Editorial judgement on UX trade-offs. "Should this be a settings page or a block toolbar control?" "Is this many sidebar controls overwhelming for the editor?" These are product decisions that depend on who the editor is and how they work. AI proposes reasonable defaults but cannot know your client's editorial team.

Brand-voice content at scale. Generating dummy content is solved. Generating actual marketing copy that matches a brand voice without a heavy human edit pass is not. The output is competent and generic, which is worse than blank for brand work.

The mental model: AI is a competent collaborator, not a replacement for understanding. You need to be able to evaluate the output — read the PHP, verify the security defaults, check the hook firing — but you no longer need to be able to write every line from scratch. The bar is "can you tell when it is wrong," not "could you have written it." That is a meaningful shift from where things stood two years ago, and it is the one most developers still under-estimate.