MCP ability pack for Oxygen 6.x + Breakdance Elements, providing tools for CSS cache regeneration, selector editing, element visibility introspection, design tab mapping, audit of compiled CSS, and backup management via UpdraftPlus.
Server has 7 tools with explicit schemas and descriptions. Naming follows verb_noun patterns (audit, backup, regenerate, merge, design-map, element-visibility). Descriptions are detailed (194 - 500+ chars) and explain WHEN to use each tool, prerequisites, and limitations, well above the 10 - 1024 baseline. However, parameter descriptions are inconsistent: some params lack descriptions entirely (e.g., 'id' in merge-css-selector), and several descriptions are in Polish, creating ambiguity for English-speaking LLMs. Output schemas are declared but minimal (type + property names only, no nested field descriptions). Error handling is present but generic (WP_Error returns). Tool composition is sound, each tool has one clear responsibility and output includes IDs for chaining. Security: no credentials exposed as parameters. Overall: strong descriptions and naming, but incomplete parameter documentation and sparse output schema detail hold the score in the C+ range.
Finds design values that were SAVED but never COMPILED — the builder's most expensive failure mode, because every tool reports success and the render simply ignores the value. Measured three separate times in one site transfer: the secondary button outline colour, the font family and weight of every button, and every border colour set on a :hover state. Each looked correct in the editor, each returned success from the write tool, and none of them reached the stylesheet. This ability compares the stored selector properties against the compiled selectors.css and reports what is missing, plus the known global-settings fields the compiler accepts and ignores. Run it after any batch of design changes, and FIRST whenever a value "does not work" even though the editor shows it set. Limitations, stated so the output can be trusted: only the base breakpoint is checked (rules inside @media are excluded on purpose), only properties whose emitted declaration name is unambiguous are mapped, and custom_css blocks count as present. A clean report means "nothing from the mapped set went missing", not "everything is perfect".
Compact map of every control an element exposes, one line per control, as "dotted.path => type". Use this INSTEAD of get-element-schemas whenever the question is "can I style this natively, and under which property path" - the full schema for one element runs 5-15k tokens and is mostly $defs and repeated breakpoint wrappers, which is why agents give up and reach for a CSS class for something the element already styles itself. Responsive controls collapse to one line marked "@breakpoint" (the five breakpoint keys are always the same). Filter with "prefix" to get just one tab, e.g. "design" or "design.typography". Pass "groups_only" to get two or three lines per element instead of every control, which is how you survey twenty elements in one call and find out WHERE a control lives before asking WHAT is in it. Read the result as: a path listed here can be written with edit-post and compiles to the element own generated CSS, so it needs no class and no selector; a path NOT listed means the element genuinely has no native control and a class is the honest answer. Read the paths verbatim: several are NOT named the way you would guess, and guessing is exactly how an agent concludes an element has no Design tab and reaches for a class instead. On EssentialElements\Div the flex/grid layout lives under design.layout_v2 (not design.layout), and padding, borders, radius and shadow live under design.container (not design.borders) - a get-element-schemas call filtered by the guessed names returns empty and looks like proof that the controls do not exist.
Parameter descriptions missing or incomplete. 'id' in merge-css-selector has no description; 'properties' lacks detail on expected structure. LLMs cannot infer parameter meaning from names alone.
Polish-language parameter descriptions in updraft/backup-now ('Zakres kopii', 'Krotka notka') create ambiguity for English-speaking LLMs. Descriptions should be in a single language matching the tool name and primary documentation.
Output schemas are minimal (type + property names only). No nested field descriptions, no indication of field types within arrays, no guidance on what each field means. LLMs cannot plan downstream tool calls without knowing output structure.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
Why an element is invisible to MCP, split by which of the three bans is responsible. Call this when get-element-schemas returns an element in "missing" or edit-post rejects an element_type. THREE SEPARATE LISTS, and they are not the same mechanism: (1) banned_by_oxygen — the hardcoded BREAKDANCE_MCP_HIDDEN_ELEMENTS constant in Oxygen's mcp/schemas/elements.php, which no site setting can change; (2) hidden_in_elements_panel — the oxygen_builder_hidden_elements WP option a human edits under Oxygen → Settings → Elements, which Oxygen pre-fills with ten Essential Elements on install; (3) restored_by_this_pack — what this pack adds back on the breakdance_builder_elements filter at priority 20. CRITICAL AND NOT OBVIOUS: list (2) DOES affect MCP. The settings page subtracts its list on the same breakdance_builder_elements filter that Oxygen runs the MCP element list through, so an element a human hides in the Elements panel disappears from get-element-slugs and get-element-schemas too. Measured 2026-08-21 on Oxygen 6.2.0-beta.5: adding EssentialElements\Button to the option took get-element-slugs from 164 to 163. So on an agent-built site: BUILD FIRST, HIDE LAST — hiding elements before the build is finished cuts the agent off from them.
Merge properties into an existing global selector instead of replacing it. The built-in insert-stylesheet REPLACES a selector's whole property map for the breakpoint it touches, so importing a rule that changes one property silently drops every other property that selector had — still true on Oxygen 6.2.0-beta.5, measured 2026-08-21: ".probe{margin:4px}" imported over a selector holding padding plus a custom_css block left only the margin. The most common source of regressions when editing an existing design system. This ability deep-merges per breakpoint: only the keys you pass change. Pass properties keyed by breakpoint id (breakpoint_base, breakpoint_tablet_landscape, breakpoint_tablet_portrait, breakpoint_phone_landscape, breakpoint_phone_portrait). Read the current shape with the built-in get-css-selectors using include_properties first. !important NO LONGER NEEDS THIS ABILITY. Oxygen fixed it in 6.2.0: mcp/design/css-to-properties.php now diverts every !important declaration straight into custom_css with the flag intact, before the typed converters see it. Measured 2026-08-21 on 6.2.0-beta.5 — insert-stylesheet with ".probe{color:#ff0000 !important;padding:8px}" stored padding as a typed property and custom_css as ":selector {\n color: #f00 !important;\n}". So write !important with the plain built-in tool. What you DO still need this ability for, besides the merge: insert-stylesheet lowercases custom property NAMES. Measured on the same version, ".probe{--myVar:4px;padding:var(--myVar)}" is stored as "--myvar: 4px; padding: var(--myVar)" — the declaration is lowercased, the reference is not, every var(--myVar) dies silently and the import still reports success. Writing custom_css through this ability bypasses that parser. Note the custom_css block compiles to a SECOND rule carrying the same selector, emitted after the typed one, so at equal specificity it also wins on order.
Recompile the builder's CSS files. Oxygen still ships no MCP tool for this — checked 2026-08-21 against the 49 tools of the 6.2.0-beta.5 server. Its own save paths do regenerate (data/save.php calls generateCacheForPost on a document save and generateCacheForGlobalSettings on a settings or selector save), so you need this only when the data changed outside those paths, or when the compiled files are simply stale. Run it after editing element trees or global settings directly, and whenever the front end does not match what the builder shows — a stale cache is the single most common cause of "the plugin update broke the site", because the compiled files can predate a plugin update and still carry its old default CSS. Pass post_ids to limit the run, omit it to do every builder post. Returns a verdict on the resulting files so you can confirm the compile actually produced output: a 0-byte post stylesheet means the run did not work for that post. Pass verbose to get the size of every file instead of the summary — on a site with 89 builder posts that listing is 178 entries and says nothing the summary does not.
Starts an UpdraftPlus backup. CALL THIS BEFORE any destructive or wide-reaching change on a site you did not build: bulk edits, a plugin or builder update, a theme migration, a database rewrite, deleting posts, selectors or global settings. TWO THINGS THAT ARE EASY TO GET WRONG. (1) This returns when the job is QUEUED, not when the backup exists. It has to work that way: firing the UpdraftPlus hook directly archives synchronously, and on a WooCommerce site with media that blew past a 120-second MCP timeout mid-backup, leaving a half-finished job — so this schedules the same hook through WP-Cron and returns immediately. Never report "backup done" off this call; poll updraft/list-backups with newer_than set to the returned started_after until an entry appears. On a store with tens of gigabytes of uploads a full backup runs for tens of minutes, so prefer what="database" when the change only touches content, selectors or settings — it finishes in seconds and is a real restore point for that class of change. (2) If the response has cron_disabled = true, NOTHING will run the job on its own and there will be no backup — say so instead of proceeding.
Lists UpdraftPlus backups, newest first: timestamp, readable date, what each one contains (baza / wtyczki / motywy / media / inne), its label, and whether it went to remote storage or sits on this server only. Two uses. (1) Confirm a job started with updraft/backup-now has actually FINISHED — an entry appears in history only once its archives are written, so a new timestamp is the completion signal; pass newer_than = started_after to poll for exactly that job. (2) Before destructive work, check whether a usable backup already exists: if the newest is days old, or contains only "baza" while you are about to touch files, take a fresh one instead of trusting it.
No pagination support declared in list-backups. 'limit' param exists but no 'offset', 'cursor', or 'total_count' in output schema. Large backup histories could exceed context window.
Error handling returns WP_Error objects with generic messages. No recovery guidance (e.g., 'Try calling X first' or 'This is retryable'). Agents cannot determine next steps from error responses.