dotsy MCP Server

Replace the whole field schema (draft; publish makes it live). Read it with get_collection_schema first and send it back edited. Collection keys exactly as list_collections returns them ("blog/" or an id). Field: { key, type, label?, help?, required?, unique?, inCard?, default? } — type text (multiline, minLength, maxLength) | number (min, max, integer) | boolean | option (options[] required) | multiOption (options[] required, max) | link | email | phone | color. Keys: letter then letters/digits/_ (≤40), not a built-in post field (title, date, excerpt, tags, image, url, slug, …; author only as text), not secret-looking. ≤60 fields per collection. settings.defaultSort: date:desc|date:asc|title:asc|slug:asc. Removing a collection's entry drops its definitions (values stay on the posts). Values live in put_post meta.custom; writes then return fieldWarnings (soft, never blocking). 400 invalid_collection_schema lists issues; 409 conflict = stale ifMatch.

Stage one content key. Body ≤60KB. Larger pages: stage_content_chunk + apply_changeset bodyStaging, dotsy push, or PUT content-raw. Edits to an existing large page: batch_patch_content.

Create or replace all site conversion goals (≤50). Required before start_test when no goal exists.

Replace the whole menu doc (ifMatch from get_menu_doc required). Each item may omit menu — the enclosing menu key in doc is used.

Partial upsert one menu item: send only the fields to change; every omitted field (parent, sort, label, kind, ref, newTab, html/childTemplate/childShell/attrs/data) keeps its stored value, and parent: null moves it to the top level. A new id needs label, kind and ref (parent defaults to null, sort to the end). To remove rich keys, use put_menu_doc. A brand-new id dedupes by (menu, ref) — reuses an existing item linking the same target instead of creating a duplicate. An id that already names an item always updates that exact item, even if another item links the same page.

Stage a blog post via structured meta+body (?format=post). Key e.g. blog/my-post.html. Declared membership: meta.collections; schedule: meta.visibleAt; SEO title override: meta.seoTitle (card always shows meta.title); custom fields: meta.custom (typed per collection by get_collection_schema; the response's fieldWarnings flag values that don't fit — soft, never blocking).

Create/replace route; contentKey derived from path if omitted. On a field that accepts null (contentKey, redirectTo, redirectStatus, preserveQuery), sending null clears it back to its default.

Create empty site; seeds brief. Globally taken slugs auto-suffix to `{slug}-{adjective}-{noun}` (default); pass onSlugTaken:"fail" to get a 409 instead of a renamed site.

Permanently delete site; confirmSlug must equal slug.

Site detail, stats, publicUrl, brief head (default-on), and readiness flags. pages[] is a ~25-entry head by default (stats.pages has the true count) — pass pages: true for the full array (capped at 1000), or enumerate everything with list_content/get_site_manifest. Large sites (200+ pages): list_pages is paginated — use cursor or get_site_manifest for the full route set.

Read-only build progress view (structure/adapt jobs or after import). structuredContent.forHumans holds the headline and progress. import_site_from_url blocks until done, so this is not needed during an import; get_site has the full state.

Content keys + hashes only (optional routes=true) — static files live in list_assets, not manifest. Diff keys+hashes before apply_changeset; stage only changed bodies.

Site shape: every layout and shared section (`sections`: fragments/* by name, with page count, whether the file exists and staged-change flag), whether a layout is missing, redirects count, menus grouped by where they show, and collections with their page placements. Use for 'is my layout missing', 'which shared sections exist and where', 'where does this menu show' — without listing every page. Read-only; indexing mirrors list_pages (a freshly imported/graduated site may still be indexing).

Update site settings: title, keyConvention, or editingMode ('live' = each editor save publishes immediately; 'staged' = saves stay drafts until publish_site). Slug is the hostname — change it with rename_site, not this tool. Set 'staged' before handing the editor to a person via start_edit_session's draftEditorUrl. signature toggles the "Made with dotsy" badge ({enabled, mark: logo|dot, reveal: pill|seal, side: right|left}); it is live immediately. trailingSlash (ignore|never|always) sets the page URL form: the other form 301s to it and canonicals/sitemap use it; live immediately. ownSeoFiles: true lets the next publish use the site's own robots.txt that blocks the whole site (Disallow: /) and its own static sitemap.xml (replaces the generated one, collection pages included); a robots.txt without Disallow: / is used anyway. aiTrain (true or false) says whether AI training crawlers may use the live site; false asks them not to, true allows them (the default). Live immediately.

Connect apex and/or www on a paid plan; starts a DNS connect session.

Read the site-wide custom code: `draft` (what the next publish puts live) and `published` (what visitors get now) for both boxes — head (before </head>) and body (before </body>) — plus `pending`, `runOn` and `maxChars`.

Current connect session and nextAction (browser URL or manual DNS guide).

List all custom hostnames and connection status for the site.

Remove one custom hostname from the site.

Set site-wide custom code: `head` goes before </head> and `body` before </body> on every page, on every layout. It is saved as a draft and goes live with the next publish_site (on a live-mode site the dashboard publishes only the code). It never runs in the editor. Up to 20,000 characters per box, stored as sent; `warnings` flags likely mistakes (unclosed <script>, <html>/<head>/<body> tags, a canonical or robots tag, GA4 already connected) without rejecting. Send only the boxes you change; "" clears a box. `runOn` ('live' = live site only, default; 'live-and-preview' = also in the editor's preview) applies at once — agents limited to drafts can't set it. Use for tag managers (GTM), Clarity, ads pixels, chat widgets and verification tags. For plain GA4, Plausible, Fathom or Umami prefer set_analytics_provider.

Propose nav wiring + blog collectionization for an imported mirror (draft, opt-in). Read-only; returns per-adapter reports + proposalEtag. Nav decline may include nextStep (e.g. mega-menu → edit fragments/header.html). Wait for mirror-vs-editable answer if requiredDecision pending.

Apply the adapt proposal (draft): wire nav into the menu store and/or rewrite the blog index to a live <dotsy-collection>. One undo covers the whole apply; a second apply is refused until adapt_site_revert (or force). Queued applies are polled to completion.

Atomically restore everything the last adapt_site_apply wrote (fragments, the menu doc when the op touched it, index, member configs, routes).

Read adapt state: proposal summary + proposalEtag (+ applied/consumed flag), undo-record presence, active or queued job progress. Use to resume in a fresh session.

Delete content key (tombstone if published). A fragments/* section that pages still place is refused with section_in_use (lists the pages) — detach it or remove it first.

Read one content key; optional selector (region) or published=true (live body). Bodies over 40KB without a selector return a truncated summary — use selector, batch_patch_content, or dotsy pull.

List content keys (HTML/md); not assets — use list_assets. Defaults to 50 keys per page — raise limit (max 1000) or pass cursor (from nextCursor) to page through the rest. Per-row `size`/`missing` are measured automatically only on pages of ≤100 rows; past that they're ABSENT (not measured, never `null`) unless you pass withSize:true, which forces measurement at a cost of one storage read per returned row. Only `size: null` + `missing: true` means the page's stored bytes are gone.

Find pages containing a phrase (grep without reading bodies). Requires do-sqlite — graduate_site first if search_unavailable.

Last structure apply job result (structuredKeys, conflicts, partial).

Flat vs structured page counts; chromeCoverage = real per-fragment page counts + missedPages (check before promising sitewide edits). Informs user — do not auto-run analyze/apply while requiredDecision pending.

Propose shared layout + fragments; persist:false first. Wait for mirror-vs-editable answer if requiredDecision pending.

Rewrite flat pages to layout + fragments (draft). Do not call while requiredDecision pending — user must pick editable first. It extracts header/footer only; report chromeCoverage counts, not 'sitewide'. Afterwards share_section with no arguments lists the other repeated blocks (banner, CTA).

Who a site's forms may email. A form's `to` address gets submissions only when it is a verified site member or it confirmed by a one-click link we email it (no account, no sign-in, no site access). action: `list` (confirmed + waiting addresses), `status` (one address: member | confirmed | pending | expired | unconfirmed), `request` (email the confirm link; 7 days, single use; max 3 per address, 20 per site and 50 per account a day), `revoke` (stop form emails to a confirmed address). A publish already sends the link once by itself when a form emails an unconfirmed address. Never ask the user for the link — it goes to that inbox only.

Read one form submission with the full field payload (explicit PII access — use only when needed).

List form submissions (metadata + bounded field preview — not full PII). Works on kv and do-sqlite backends.

Remove one item from a menu (page stays live).

Read a menu's draft items and etag; menu rules → get_skill('nav').

Read the whole menu doc (all menus + shell/template), etag for put_menu_doc, and issues (by item id) — page_not_routable, page_missing, page_dead_link, page_is_redirect — computed from the draft routes, so a link that won't show or leads nowhere shows up before publish.

Re-key a page by slug (draft or published static). Server-side hash copy — no body round-trip. Default deletes static routes (obscure-URL); movePath:true repoints the route, rewrites menu refs and other redirects aimed at the old path, rewrites root-relative and https://{slug}.dotsy.site/ hrefs in draft HTML (always linksRewritten; not custom-domain absolutes or script/code samples; linksRewriteTruncated when scan caps at 100 HTML keys (500 on sites with full content search); publish to ship), and — only for an already-PUBLISHED page — leaves a 301 there too (leaveRedirect, default true; no effect on a never-published page, or when movePath is false). Retry-safe: a part-way failure (409 rename_failed) rolls everything back, so resend the same call. On fragments/<name>.html it also rewrites every <dotsy-fragment name> reference and hide marker in the same all-or-nothing commit. Then publish_site.

Change a page's URL: moves its route from `from` to `to`. Menu links and other redirects that pointed at the old path retarget automatically; a 301 is left at the old path by default (leaveRedirect). Root-relative and https://{slug}.dotsy.site/ hrefs in draft page, layout and fragment HTML are rewritten (always linksRewritten; not custom-domain absolutes or script/code samples; linksRewriteTruncated when scan caps at 100 HTML keys (500 on sites with full content search); publish to ship). Retry-safe (resend the same call after any error). Errors: 404 route_not_found (from isn't a route); 400 reserved_route (to is under /__dotsy/) or invalid_body (missing/malformed from or to); 422 rename_same_path (from equals to already), rename_pattern_route (from or to contains a :param segment — pattern routes can't be renamed this way), rename_redirect_route (from is itself a different redirect — edit its redirectTo instead); 409 rename_experiment_running (an A/B test is running on from — end it first) or rename_target_exists (to is already a different route — unless to already serves the exact content from would, which is treated as already done and returns 200). Then publish_site.

Rename the site slug (public hostname {slug}.dotsy.site). Owner-only. After success: use the returned slug in every later tool. Published site: the old slug redirects on the API (301 for reads, 308 for writes) while it stays reserved; the old host 301s and stays reserved until delete or rename-back. Never published: the old slug 404s on the API and the old hostname is released (no 301; create_site can reuse it). Analytics history stays on the old slug. HTML absolute URLs are not rewritten — publish_site refreshes sitemap; find_replace if stored HTML must change.

Open a draft session. Returns preview/edit URLs for the person (see linkGuidance). draftEditorUrl only when the site gives AI agents Full access. path defaults to '/'.

Start one draft site with no dotsy account. Returns slug + guestToken — pass guestToken on every later tool call until claim_site.

Start a same-URL A/B test (draft or live source). Rules → get_skill('verify') § A/B.

Read-only traffic overview (default) or full breakdowns via include[]; optional bot id for crawler page drill-down.

Connect a third-party analytics provider (GA4, Plausible, Fathom, Umami) or remove it (provider 'none'). dotsy adds the provider's script to the <head> of every live page — live immediately, no publish needed. Replaces any earlier provider (one per site). Use this instead of pasting a GA4 gtag snippet into a layout. Other tags (Google Tag Manager, ads pixels, Clarity, search-console verification <meta>) go in set_custom_code.

Delete a static asset by key. The editor and draft preview drop it at once; the published site drops it at the next publish.

Read one asset (metadata + base64 + stored sha256). sha256 is the stored blob; may differ from upload bodySha256 when IMAGES re-encodes rasters. Uploaded files are live in the editor and draft preview at once; the published site picks them up at the next publish.

Upload 1–200 CMS assets; key = path after /assets/, or sitePath for an exact URL path (e.g. /.well-known/security.txt, /ads.txt). Each item: exactly one of bodyBase64, bodyText (plain text for .css/.js/.mjs/.json/.svg/.txt/.xml only — not .html pages), or importFromUrl. Optional bodySha256 verifies submitted bytes before write; response sha256 is stored bytes (JS/CSS match; WebP rasters may differ). JPEG/PNG/static GIF are re-encoded to WebP (response key photo.webp; /assets/photo.jpg keeps serving it) — use the response key in CSS/JS; optimize:false stores the bytes as sent at your key; animated GIFs are always kept. Uploaded files are live in the editor and draft preview at once; the published site picks them up at the next publish.

Stage many content keys (pages, layouts, fragments). Content only — not routes or assets. Each body ≤60KB and the whole call ≤80KB; larger pages use stage_content_chunk, dotsy push, or PUT content-raw. May queue above the inline threshold.

Read the site's collection field definitions (`_data/collections.json`, draft view): { schema: { v: 1, collections: { "<key>": { fields[], settings? } } }, hash }. hash is null when no schema exists yet — pass it as put_collection_schema ifMatch. 409 collection_schema_invalid { hash, issues } means the stored file is broken: fix it with put_collection_schema using that hash.

Rename a collection's display name and/or noun (e.g. the generic 'posts' key shown as 'Recipes' / 'recipe') — cosmetic only, stored separately from the tag that defines membership; doesn't need publish. Provide at least one of name/noun. null (or "") resets that field back to its derived default — different from omitting it, which leaves the current value alone.

Remove a draft A/B test (releases variant bodies; does not publish).

Pending draft changes: content keys and routes with status. May include requiredDecision (publish_confirm) before publish_site.

Edit one content key via ops[] (selector ops or replace-string; wrap/unwrap/move restructure siblings). Stages draft; op shapes in schema.

Partial route update (only sent fields change; an OMITTED field is left alone). On a field that accepts null (contentKey, redirectTo, redirectStatus, preserveQuery), sending null CLEARS it back to its default — different from omitting it. To set redirectTo/redirectStatus, include purpose: "redirect" in the SAME call — the check looks only at what's sent, not the route's stored purpose, so a PATCH that sets redirectTo without repeating purpose: "redirect" (even on a route that's already a redirect) 400s invalid_body.

Your dotsy bug reports. Omit id for the list; pass id for one ticket and its public replies.

Report a dotsy bug, or ask a support question (kind question — AI answer from the help centre). Bugs attach recent errors and return a ticket id.

Remove a route by path.

Get one route by path (home is '/').

List voters on one board item (voter id + source only — no email).

Remove one voter's vote and drop their release subscription for that item.

Account, user, memberships[]. Pass workspace on writes when memberships.length > 1.

List site activity or fetch one change diff (pass id). Filters: who, state, actor (me), since, limit≤20. Notes from context are visible to everyone on the site — no secrets.

Stage writes[] + deletes[] in one call; attach routes via writes[].route. writes[].ifMatch (hash from get_site_manifest; "" = must not exist) guards each key (a delete takes { key, ifMatch } too); atomic:true makes one stale key write nothing (409 changeset_conflict). At most 10 writes; each body ≤60KB and the call ≤80KB. Larger new pages: stage_content_chunk then writes[].bodyStaging, or dotsy push / PUT content-raw.

CMS uploads under assets/ only — not mirrored clone CDN (get_site mirroredAssets or import warmJob). Not content — use list_content. Uploaded files are live in the editor and draft preview at once; the published site picks them up at the next publish.

patch_content on multiple keys; all-or-nothing validation. Preferred edit path for pages over 60KB. replace-string value/replace ≤60KB; whole call ≤80KB.

List board items with status, stamp fields, and baked vote counts (draft overlay included). Live vote totals: site host GET /__dotsy/votes.

Update `_intent/site.md` (site brief) or a page brief when `page` is set. See field descriptions for rules, notes, and head enums.

Move the guest draft onto your signed-in dotsy account. Requires OAuth Connect; pass the guestToken from start_site.

Collections on the site — name, noun, item count (or 'not published yet'), URL pattern, and whether any member is staged. Includes a collection only a draft page lists. Returns a bare array (not wrapped in an object) — pass each row's key verbatim to inspect_collection/update_collection.

Give one page its own copy of a shared section; every other page keeps the shared one. A draft — publish_site to go live. A section the layout places is hidden on that page instead (share_section name + hide).

Discard unpublished drafts: all (omit key/keys/routes/menus), one content key (key), menu draft only (menus:true), or a selection (keys, routes). Destructive — check get_draft_status first.

Open the interactive page canvas to change content: the user clicks an element, then you call patch_content. Use when the intent is editing — not for verify-only checks (use preview_page).

Estimate days to reach promote readiness for a page (traffic baseline). Read-only.

Record what was learned from an ended A/B test (≤500 chars; null clears).

List site conversion goals (pageview paths and form submits). Required before start_test.

Enable site search & collections (kv → do-sqlite), one-way. Unlocks search_content, collection indexes, visitor /__paige/search, and moves any R2 form-submission ledger into SiteDb. 409 blocked covers three unrelated gates: an unsupported source backend (no override, ever); strictDrafts — default true, this client sends true unless you pass strictDrafts:false — blocking on unpublished drafts or unrecovered KV content rows; or a KV snapshot that looks incomplete (e.g. a page deleted in the last couple of minutes). force:true overrides ONLY that last, incomplete-snapshot gate — never the backend or strictDrafts gates. The blocked response's `reason` names which gate fired, and `audit` (+ `audit.warnings`) is aggregate COUNTS only (published/route/pending-draft counts) — never row keys — so judge the size of the gap, not a row list, before forcing. The flip is one-way: anything outside the frozen snapshot is unreachable once it lands.

Add an idea from outside the site (draft board item or inbox suggestion). Idempotent on externalRef.

Search draft/published board items and open suggestions by title token overlap.

Import URL to draft. On complete, response may include requiredDecision — preview then ask mirror vs editable before structure_*.

Suggest fields for one collection from a sample of its items' meta.custom values (published data, else drafts) — writes nothing. Returns { sampled, suggestions: [{ field: { key, type, options? }, count, samples, defined }], skipped }. Review, then add the ones you want via put_collection_schema.

Collection state — count, server.mode, sample entries, tag (membershipKind, match=) when discovered; plus name, noun, URL pattern, layout, where it's listed, and post fields used. Pass status/q/cursor/limit (any one, even empty) for a full items[] page instead: status draft = never published; a published post with unpublished edits is status published with hasChanges:true (never draft) and shows its staged title/date/tags/author; scheduled = future visibleAt. Never-published drafts come from staged content (no extra reads, no publish needed to see them); published/scheduled come from the index. limit defaults to 50, clamped 1-200. Omit all four for just the base inspect (no items[]).

Add a roadmap, changelog and/or help centre to a site as drafts (never overwrites; rerun finishes a partial install).

Change a roadmap/board item status (stamps dates; on release drafts the changelog post and links both ways). Never hand-edit *At, daysToRelease, or votes.

Mark Update notifications read (ids or all). Needs-you Actions refuse with 422 action_not_markable — fix the cause (it clears itself); only the person can dismiss them in the dashboard.

Decide one form suggestion from list_suggestions: approve (new draft board item; stages content, does not publish), merge (into an existing board item, subscribes the voter/email) or decline (no board item).

Read-only scaffold for a new post — suggested key/path, membership meta, blank or duplicate body. Then put_post (+ optional put_route).

Account notification inbox (Needs you · Updates). Summary-first: tab counts plus compact rows with nextAction (MCP tool or dashboard URL). OAuth person only — not API keys.

Shrink heavy images in place (0.98): JPEG/PNG/still GIF → WebP at the same address, so no HTML changes. keys = site paths from verify_page health.assets[].ref — /assets/… uploads or clone paths like /wp-content/… (1–50 per call; batch keys from several entries). Skips animated images, already-small files and non-images (`skipped`); idempotent. The replaced file is kept under .originals (never served). Spends the daily image budget like an upload — dryRun:true reports the saving without spending it. The draft serves the smaller file at once. Your live site shows the smaller file after you publish (some imported files sooner).

List pages on a site — route settings (path/title/navLabel/purpose) by default, one row per route. Renders a page-picker for humans (redirect and worker routes left out there — nothing to open in edit_page); agent reads structuredContent. Use after picking a site — not for editing (use edit_page) or verify-only checks (use preview_page). Pass view:'pages' to group by layout instead — adds layout/noLayout/q filters, what each page places, and indexing progress; use that for 'what pages use this layout' or 'is this page indexed yet', not for route settings. Paginated — pass cursor (from nextCursor) for more; partial:true means this page is a window, not the full route set.

Read-only workspace plan and pooled usage; pass slug for this site's add-ons. Call before a large build.

Plan check before publish or large build (counts, plan page cap, seo suggestions; imports: form_unwired advisory, webflow_store blocking). When ok is false, blocking lists the failed suggestion codes. Read-only.

Verify-only composed HTML check; home preview may include requiredDecision — ask mirror vs editable before structure_*.

Set the canonical host (301 target) when multiple hostnames are connected.

Promote a winning A/B variant to live — explicit user instruction only; see get_skill('verify') § A/B.

Publish draft to live. handoff required; feedback optional. May return advisory form_unwired / store warnings (never blocks) — tell the user.

Re-apply DNS with stored Cloudflare OAuth grant (no browser).

Poll one hostname until DNS/cert propagates.

Build/finish collection indexes after publish returns collectionDataPending or collectionSqlPending.

Set final menu order. Pass ifMatch from get_menu; stale → 409 with current items.

Case-insensitive find→replace across pages in one staged commit. Batch all replacements into ONE call — never parallelize find_replace on the same key (concurrent edits surface as per-key conflict). find/replace ≤60KB.

Add information or answer dotsy's question on your own report.

Bring an earlier published version back as unpublished changes (page content only — routes, menus, images and settings stay). version: a number from list_versions, or "previous". dryRun: show what would change. Then review and publish_site, or discard_drafts to undo; publish:true makes it live at once.

Undo one activity event by staging the prior version as draft only — call publish_site to go live. Respects AI access.

Review queue: what changed and isn't live yet. list (counts first) → get one item (hidden changes, optional diff, bodyChanged: false means only title/meta/route fields changed — review hidden, not the visible page) → mark_ready / unmark (≤50 ids). Then publish_site onlyReady:true. Drafts-only sites: a person marks.

Make a block that repeats on several pages one shared section, or change where one shows. No arguments: lists blocks repeated on 3+ pages. sourceKey + selector (or selectors: a sibling run, one wrapper): a dry run finds every page with it; dryRun:false shares it. name + scope/hide/show: move an existing section to/from every page, or hide/show it per page. Over the per-call cap (too_many_pages, nothing written): share with keys, then again under the same name from a page that still has a copy; for every page then move with name + scope every_page + keys (max-1 pages a call; put all hide on the first) until all are moved. Always a draft (publish_site to go live; undo: revert_change).

List your dotsy sites. Renders a site-picker for humans; agent reads structuredContent. Use to choose a site — not for editing content (use edit_page) or verify-only checks (use preview_page). Defaults to your active workspace; pass workspace (org id or slug from get_account_context memberships) to list another workspace's sites without switching the dashboard session.

Skill body + optional questions[]. Accepts an id OR a synonym (e.g. 'list', 'menu') — resolves aliases and suggests near-misses. Start with intake when ambiguous.

Paige build skills (id + whenToUse). Pull get_skill on demand — do not duplicate in tool text.

Append one ≤60KB chunk of a large new page. Omit stagingId on seq 0; pass the returned stagingId on later seq. Set finalize:true on the last chunk, then apply_changeset with writes[].bodyStaging.

Optional mid-session feedback (required feedback rides with publish_site).

List form suggestions awaiting moderation (duplicate hints against the board).

Take back the last publish at once: restores the version before it for the pages that publish changed and publishes them, and the undone changes stay as unpublished changes (`restaged`). On a live-mode site nothing is kept (its next save would publish it again): the answer has `redo.version` — restore_version with that version and publish: true brings it back. publishVersion = the version publish_site returned. Fails with not_latest if another publish landed since. If its publish fails, the undo stays as unpublished changes — publish_site finishes it.

Pre-publish lint (tags, assets, layout, SEO, speed and accessibility). Summary first: health issues grouped by code, each with fix:{tool,args,fill?} — supply the fill args, then call. health.summary counts are whole-site (groups/codes filter only issues and assets). Pair with preview_page — see get_skill('verify').

List the site's published versions, newest first (who, when, pages, live, restorable). Use before restore_version.

Poll one hostname until it is connected or the timeout is reached.