eden_add_table_rows
ChatGPTAppend rows to an existing table (including recurring ones — see repeat). Read it first with eden_read_table for exact column names and options. Unknown select options auto-create; anything unresolvable comes back in warnings — fix and retry those rows only. Rows are lightweight records; no documents are created.
eden_analyze_creator
ChatGPTUSE WHEN the user asks about a SPECIFIC creator's content by name, handle, or URL — "best tweets from @dankoe", "top videos from Ali Abdaal this quarter", "what does Naval post about". This is the right tool EVEN IF the user has saved that creator's posts to their workspace — it ranks against the full indexed corpus and the creator's own baseline, not the hand-picked swipe file (see server instructions for routing). WHEN THE CREATOR IS THE USER THEMSELVES asking about their own performance ("how did my posts do", "my top posts", "my engagement"), prefer eden_get_analytics / eden_list_analytics_posts — the analytics warehouse has their PRIVATE metrics (impressions, saves, watch time, email opens) plus period deltas that this public index can't see. Use this tool on their own handle only when analytics isn't set up, or when they explicitly want the public-index view of their account (outlier baseline, how their posts rank publicly). Returns a profile-level analysis: identity, totals (avg views, engagement, post count), topic/format breakdowns, and a curated set of ~8 best posts (real outliers >=1.2x the creator's own baseline, padded with recent posts when the outlier shelf is thin). The response includes a pre-rendered text field (markdown rollup) you can paraphrase from directly, plus structured topPosts[] entries each carrying a contentId for follow-up eden_read_social_post calls when you need the full body or transcript. FIRST-TIME CREATORS: when nobody on Eden has searched this creator before, the tool auto-pulls their recent posts inline (waits up to ~25s — never immediately re-call it). It then returns either ok: true with a notice, or ok: false, status: "indexing" with recentPosts, etaLabel, estimatedTotalPosts, progress, and a partial creator card. Then: render recentPosts immediately when present (as "what they've been posting lately", not "top posts"), show the partial card (bio, followers) so the reply isn't empty, prefer "we have N of ~M posts" from progress over etaLabel when progress.synced > 0, and use plain language with the user (never "backfill" / "sync" / "ingest" / "indexing").
eden_cancel_scheduled_post
ChatGPTCancel a scheduled post or delete a draft in Eden by id, removing it from the publish queue. Use this to clean up a mistake or a throwaway draft. Find the id with eden_list_scheduled_posts. This cannot cancel a post that is already publishing (it returns a conflict) or one already posted. The cancel is permanent.
eden_connect_social_accounts
ChatGPTCheck and set up the user's social account connections in Eden. Three actions: status lists what is connected right now (per platform, with handles; Substack includes its publication address). get-link mints a secure, personal account-linking link (an Eden-branded page) — give it to the user to open, where they click each network (X, LinkedIn, Instagram, Threads, TikTok) and authorize it. sync registers what they linked — ALWAYS call it after the user says they finished linking, or the new connections will not appear in Eden. Substack is the exception: it cannot be linked through this page; the user connects it in the Eden web app under Settings → Social accounts (https://app.eden.so/?settings=scheduling), ideally with the Eden browser extension or desktop app signed in to Substack. The linking link is tied to this user's workspace — never share it anywhere public.
eden_create_board
ChatGPTCreate a new empty board (a canvas) in the user's Eden workspace and pin it to the top of their sidebar. Use this when content needs a home that doesn't exist yet — e.g. before saving links, posts, or a note, when the user names a board you can't find with eden_search_workspace_items. Returns the board's itemId as boardId, which you pass to eden_create_note, eden_save_links_to_board, or eden_save_posts_to_board. Search for an existing board by title first with distinctive tokens ("rock bottom", not only "in rock bottom" / "my hooks board"); only create when there's genuinely no match. Do NOT call this in the same tool batch as eden_search_workspace_items — wait for search results before creating. Server-side, same-titled boards are reused when found (reused: true); different titles still create distinct boards. When the result says reused: true and the user asked for a NEW or SEPARATE board, STOP before saving anything to it — that board already holds their content; tell them it exists and ask whether to use it or create one under a different name. FOLDERS: sidebar folders are visual organization for BOARDS only — items can NEVER live in a folder, only on a board. So "put these in my <folder> folder" always means a BOARD inside that folder: find the folder first (eden_search_workspace_items returns matching sidebar folders in folders, with the boards each holds), then pick one of its boards or create one here with folder set — never a parallel board pretending to be the folder. folder places the board (new OR reused — reuse + folder is how an existing board moves into a folder) inside the named folder, creating the folder when none matches. When the user names a folder but no board in it fits and none is implied, ask which board they want (or offer to create one) rather than guessing.
eden_create_note
ChatGPTCreate a new markdown item in the user's Eden workspace. TWO PRESENTATIONS: presentation: "document" (default — the UI calls these "Documents") for drafted content the user keeps editing (post drafts, research summaries, outlines); presentation: "card" for a short canvas-visible text Card (the UI presents it like a sticky) — use card for "capture this", ideas, reminders, labels, quotes, and short commentary. Choose Library for a workspace-only item or a board to add one placement of the same canonical item. On a board, section puts it inside one of the board's existing sections (ids/titles from eden_read_board). content is the full markdown body (required for cards). If title is omitted on a document, Eden derives it from the content's first heading/line; cards take no title. Cards take no color: they always use Eden's default card color, so they switch with light and dark mode. Returns the created itemId, which works with eden_get_note_markdown and the eden://workspaces/{workspaceId}/notes/{itemId} resource on later turns. Write the body WITHOUT blank lines between blocks — Eden's editor already adds vertical spacing per block, so blank lines render as double gaps. Blank lines between blocks are stripped on save (blank lines inside fenced code blocks are kept). This tool only CREATES new items — editing goes through eden_update_note. To save an existing URL or indexed social post as a card, use eden_save_links_to_board / eden_save_posts_to_board instead (see server instructions). Retrying an identical call is safe and lands ONE item; changing any argument creates a new one. Notes are for PROSE. A structured collection — many entries sharing the same fields the user would sort, group, or check off — should be eden_create_table, never a note whose body is mostly a markdown table. When the user asked for a table, tracker, calendar, or bank, call eden_create_table ALONE — a document nobody asked for is the wrong answer. Write a note as well only when the user's own ask carried the thinking (a season plan, a strategy plus its calendar): say in one line that you are writing it up as a doc, proceed without waiting, then call eden_create_table with embedInNoteId to drop the live table into it where the records belong.
eden_create_table
ChatGPTCreate a TABLE item (a database-style list: to-do list, project tracker, content calendar, hook bank) with columns, seed rows, and a view in one call. Pick REAL column types, never all-text. A column you plan to GROUP BY must be a select (weeks, phases, sprints, tiers: select with options like 'Week 1'…'Week 12', never a bare number column — only select/rating/checkbox/date columns can form groups; a number/text groupBy target is auto-converted to a select built from its values, but ask for the select directly). For every select/multiSelect option, provide an explicit intentional color in columns.options; do not rely on fallback rotation. view.groupBy sets the grouping ('done' for todo lists); view.layout 'list' = a to-do list (one checkable line per row — pair it with a date groupBy for Overdue/Today/Tomorrow buckets), 'board' = kanban, 'calendar' = month grid. Read it back with eden_read_table; fix anything in warnings with eden_update_table / eden_update_table_rows instead of recreating, and when a warning touches the view or grouping, read the table back and VERIFY the groups hold rows before telling the user it is done. Warnings are plumbing addressed to YOU, not to the user: act on one silently and never quote, paraphrase, or narrate it in your reply. WHAT BELONGS IN A TABLE: repeating records that share the same fields — posts, hooks, tasks, links, contacts, experiments — anything the user would sort, filter, group, check off, or add another of. A paragraph, or one or two facts with no repeating shape, is not a table. Design the columns around the row's NAME: the name holds the thing itself and every column holds a different fact about it, never a restatement of the name (full rule on rows.title and columns.name — read them before choosing columns). BUILD THE TABLE, NOT A DOCUMENT: a table ask gets a table, always — never wrap it in a note the user did not ask for. Three cases: (1) the turn also produced real strategy the table cannot hold → build the table, then OFFER the doc in one line at the end ("Want me to write this up as a doc with the table inside?") — if you wrote substantial strategy or reasoning into THIS reply and it is not in the table, that IS case (1): having already spent the thinking in prose is exactly what the offer is for, never a reason to skip it; (2) go straight to building a doc only when the USER'S OWN ask carried the thinking — "plan out my content season and give me the calendar" does, "make me a table of hooks" does not; (3) on that straight-to-doc route, say so in one line first and then proceed — never stop and wait for a yes. AFTER BUILDING, tell the user in ONE line the two judgment calls you made: what went in the Name column, and how it is grouped — e.g. "Named each row by its hook, grouped by Tier." (in case (1) the whole close is that line plus the offer: "Named each row by its hook, grouped by Tier. Want me to write this up as a doc with the table inside?"). EXACTLY ONE thing may ever follow that line, and only in case (1): the doc offer, which is REQUIRED there — so this line is the end of the reply only when there is no strategy left to offer. Nothing else may follow it or join it: the table already renders its column list, so never restate or explain the columns you added ("I also added a Status column so you can track…") — that is padding. END-OF-REPLY CHECK, every table turn: reread the reply — if ANY analysis, strategy, or reasoning prose stands in it beyond that one line, you are in case (1) and the reply MUST end with the doc offer as a literal question ("Want me to write this up as a doc with the table inside?"); no other closing offer — drafting the posts, suggesting next steps — can stand in for it, and ending such a reply without that question is a mistake. EMBED IT IN A NOTE: embedInNoteId renders the table as a live block inside that note — the way to answer "pull these posts out of my note into a table". replaceInNoteText is the passage the table stands in for (the bullet list or markdown pipe table it was built from), copied from the not…
eden_get_analytics
ChatGPTThe user's OWN social analytics across every connected platform (X, LinkedIn, Instagram, TikTok, YouTube, Threads, Facebook, Substack): period totals with vs-previous-period deltas, follower counts per account, current OUTLIER posts (posts beating the creator's own baseline, with links), over-performing topics and formats, and benchmarks vs an automatically matched peer cohort. This is their private performance data from Eden's analytics warehouse, NOT public creator research — for other creators' posts use the research tools. Returns a compact digest sized for context. The response's window echoes requestedDays vs the days actually served (history is clamped to the workspace tier's entitlement) — mention the clamp if they differ. First call on an untracked workspace AUTO-STARTS tracking when accounts are connected. Recovery states (activating: true, accountsConnected: false, analytics unavailable) each come back with an explanatory note plus nextStep / fallbackHint — follow those and the server instructions; never present zeroed or still-importing totals as real numbers. A youtubeDelay note means a newly connected YouTube channel is waiting on YouTube's shared daily data limit — relay it as a temporary delay, never as the channel being ignored or broken. whatsWorking ranks the user's hooks and post structures over its own days window: multiple is a typical post in that group against their typical post (1.4 = 40% better). For per-post rows (charts, dashboards) use eden_list_analytics_posts instead.
eden_get_plan
ChatGPTRead the user's plan (their Planner): a kanban of post ideas. Returns stages (in order, each with name and count; each planner names its own, and role marks its new-idea, review, scheduled, and done stage), cards (id, text, stage, date, pillar, format, platforms, topic, angle, notes preview, linkedItemCount; on a planner, research and draft name the board or doc in its Research slot and the doc in its Draft slot (id, title, type); a main card such as a long video lists the pieces cut from it in pieceIds, and each piece has parentId and partOf, its main card's line; on a planner, posts lists the card's Publish posts with status, time, platforms, and first line), total, truncated, and weekdays (which pillar runs which weekday, e.g. { mon: "Focus" }). Dates are the user's local days, YYYY-MM-DD; null = unscheduled. Filter by stage, pillar, a from/to day range, unscheduled, or parent (a main card's id, for just its pieces). Archived cards (put away by hand, or a week untouched in the posted column) are left out and counted in archivedCount; archived: "include" mixes them in and "only" reads just them (each marked archived: true). Read it before eden_update_plan to get card ids, column names, and open days, and to avoid adding an idea that's already on the board. Planners: when the result has planner, the user's plan is a table in their Library (their plan tables are in planners: planners and any table with a Posts column, e.g. one per platform). Every table has its own columns and stage names (it can be anything), so go by what the result says, never by a template's names. Reads and writes default to the FIRST plan table; pass planner (a table's name or id) when the user names another table, and ask which one when they have several and it isn't clear. With no plan table, their old plan board moves onto a "Content planner" table on its own (once; old cards stay stored), or adding cards makes a fresh one (the note says which; pass it on). On a planner, stage is one of its stage column's options (stages, each with a role when it has one: new is where a new idea lands, review, scheduled, done; the others are working stages), date is its Date column, platform is its Platform column's choice, and parent isn't supported (pieces become their own rows with a note). archived: true takes a card off the planner, the same as Remove from Table in the app, only on the user's ask; archived: false brings it back. Those cards are left out of reads and counted in archivedCount, and reading with archived set to only lists them so you can restore one. An idea already on the planner (same words) is skipped, not added twice: the result lists it in skipped. A card's posts are its Publish posts (id, status, scheduledFor, platforms, preview). To post a planned idea, schedule or draft it with planCardId (the card id): the post joins the card's Posts column and the card moves itself (to whenScheduled, then to its done stage once it's out), so never move a card with posts by hand. A card's research (a board or doc) and draft (a doc) are the work it's made from: set them with the id you got making the board or doc (null clears), and open draft before writing a new doc for a card that has one. A column the planner doesn't have is skipped with a warning: pass that on only if it matters.
eden_get_seller_overview
ChatGPTRead-only: the user's Eden store in one call. Their store handle, every product with page views, buyers, active buyers this week, and gross and net revenue (minor units of the store's currency: cents for USD), the last 20 sales (amounts before tax), their creator page, and referral earnings. Use it for questions like "how are my sales this week" or "which product sells best". status: "not-enabled" means this account has no store; say so plainly. Products, prices, and sales are changed in Eden, not here.
eden_get_strategy
ChatGPTRead the user's content strategy from the Strategy tab in Eden: who they're for (audience), their story, mission, beliefs, enemy, who their audience becomes, who shaped them (shapedBy), their platforms, and their pillars. Each pillar has a lens (how they come at it, e.g. "Through psychology"; absent on older pillars), a claim, the weekdays it runs on their Planner, and topics with proof from real breakout posts in Eden's index (null = added by hand) and angle names. deeper (null until they do the Go deeper part of the page) holds what they're against, why now, where they take their reader, their role, who the reader becomes, the rules their reader lives by, and their own words (names with meanings): write in that voice and use their words where they fit. Pass pillar (its id or name) to get that one pillar with every angle's post ideas; an idea with onPlanner: true is already on one of their planners, so don't pitch it as new. Read it BEFORE writing content ideas, planning posts, or answering "what should I post", and keep new ideas inside a pillar. strategy: null means they have not made one yet (status "generating" = it's being made right now). They make one on the Strategy page in Eden (https://app.eden.so). The posting calendar is eden_get_plan.
eden_import_scheduling_media
ChatGPTHost an image, video, or PDF for a scheduled post from a link, without sending its bytes through the conversation. Pass file (a file from the chat, in clients that provide one) or a public https url. Eden downloads it (max 512 MB), hosts it, and returns a ready asset for the media array of eden_schedule_post or eden_publish_post_now. Supported types: image/jpeg, image/png, image/webp, image/gif, application/pdf, video/mp4, video/quicktime, video/webm. For a local file with no link, use eden_upload_scheduling_media instead.
eden_list_analytics_posts
ChatGPTPer-post rows from the user's OWN analytics warehouse — the raw material for charts, dashboards, and 'which posts did X' questions. Each row carries platform, posted date, link, text preview, thumbnail, content type, topics, the metrics Eden has synced (views/likes/comments/shares/saves/impressions/reach/watch time/engagement rate, Substack email opens, tracked link clicks), and index-derived signals (outlierScore = multiple of the creator's own baseline, creatorPercentile). Threads roll up into their root post with threadTotals. Null metrics are omitted; a metricsStatus of 'syncing' means numbers are still filling in (first import takes a few minutes — do not read missing metrics as zeros). Sort by any metric to get top posts. For the summary/insight layer use eden_get_analytics.
eden_list_scheduled_posts
ChatGPTList scheduled-post rows for the workspace or one schedule: drafts, scheduled, publishing, posted, partial, failed, or cancelled. Use this to inspect the queue, confirm what was just scheduled, find a post id for a later edit/cancel tool, or read recent source posts before repurposing them into another schedule/platform set. A Substack post already handed to Substack's own scheduler reads scheduled with scheduledOnSubstack: true and goesLiveAt on its target until it goes live. A TikTok target sent to TikTok drafts reads posted with inTikTokDrafts: true and no link until the user posts it from the TikTok app.
eden_list_schedules
ChatGPTList the user's Eden posting schedules (social sets, grouped under brands), active connected accounts, text limits, timezone, recurring posting slots, and upcoming free queue slots. Use this before scheduling when you need a schedule id, connected platform list, timezone, or next open slot, and before eden_update_schedule to read the current slot times. MCP scheduling supports text-first posts for X, Threads, LinkedIn, and Substack notes; Instagram and TikTok require hosted media in the write tools. Substack accounts include captured publication tag names when available; reuse those exact names for article tags. Ghost accounts (one per site) include ghostNewsletters to offer when emailing.
eden_list_workspaces
ChatGPTList the Eden workspaces the authenticated user belongs to, with id, name, slug, and role. A client workspace (agencies keep one per client) carries billingParentWorkspaceId, the agency workspace it belongs to; match a client the user names ("save this to Acme") by name. Call this first when the user hasn't told you which workspace they want; the returned id is what every workspace-scoped tool needs as workspaceId. Tool routing from there is covered by the server instructions.
eden_publish_post_now
ChatGPTImmediately queue a social post for publishing in Eden. This is a real publish action, not a draft or proposal. Use only when the user explicitly asks to publish/post/send now. Composing: preserve the user's line breaks exactly in text; a media-only post may use an empty text. Text-only works on X (twitter), Threads, LinkedIn, and Substack notes. Instagram and TikTok need hosted media (TikTok also takes a photo or a photo carousel, up to 35 images). media takes already-hosted public URLs from the scheduling media upload tools, never local paths. Shared images go to EVERY selected platform, LinkedIn included; never strip them from one unless the user asked. A PDF on linkedin publishes as a document carousel. Media for SOME platforms only: mark the asset scheduleMediaRole: "platformScoped" AND list its id in each intended perPlatform.<p>.mediaIds; perPlatform.<p>.mediaIds: [] makes one platform text-only. YouTube is Shorts only: one vertical video plus perPlatform.youtube.title. X / Threads threads: pass segments (too-long text without them is auto-split). X: text that is ONLY another post's x.com URL publishes as a native repost (needs X advanced automations; the response warns when they're off); commentary plus a post URL is a quote post. Long-form: article with explicit platforms from twitter / substack. Before targeting Substack, ask for audience, whether to send email, and whether to add a paywall; never infer them.
eden_read_social_post
ChatGPTRead the full body (and transcript / carousel slide text, when available) of a single social post — by DB id (contentId + platform from a prior social tool result) OR by url (a saved link card, a pasted x.com / youtube.com / instagram.com / tiktok.com / linkedin.com / substack.com URL, or a loom.com video — hydrated server-side on a cache miss). Pass ONE of the two. Use after eden_analyze_creator or eden_search_social_content when you want to quote specific prose: those return ~400-500 char previews, this returns up to 24K chars of body and 180K chars of transcript. Transcripts only exist for YouTube / Instagram / TikTok; Instagram/TikTok PHOTO CAROUSELS return the slide text as slideText (the caption alone is usually just a promo blurb), and Instagram reels / TikTok videos return the verbatim captions and overlays burned into the video as onScreenText (the transcript only covers audio). Uncached transcripts/slide text/on-screen text trigger an on-demand vendor fetch (~1-8s, costs vendor quota) — set attemptLiveFetch: false when sweeping many posts. URL mode: unsupported URLs return status: "unsupported" (use your client's web fetch); a post still hydrating returns status: "queued" — ask the user to retry in a moment. NOT for Eden's own board links (.../spaces/<id>) — those go directly to eden_read_board.
eden_read_table
ChatGPTRead a TABLE item (Eden's database-style list, type === "table" in eden_list_workspace_items / eden_search_workspace_items): column schema plus each row's cell values (first 2000 rows — the route's cap) — the right tool for "what's in my content calendar" or any rows-and-columns question. Returns properties (columns with option lists), rows (title, done, a values map keyed by PROPERTY ID — join through properties; select cells hold an option id, dates are epoch ms, rev is the row's edit stamp to echo back as expectedRev when writing), referencedItems (titles for item-reference cells), and viewConfig. Rows with isRecord: false are real documents readable via eden_get_note_markdown; rows carry itemIds for the table write tools. A posts cell (a planner's Posts column) holds Publish post ids; postSummaries maps each id to its status, time, platforms, and first line. Posts join that cell when scheduled with planCardId on eden_schedule_post; it can't be written here. person cells read as member names; people lists who can be assigned. THE USER'S VIEW: viewSummary (when present) names the LAYOUT they are looking at (list = a to-do list, one checkable line per row; board = kanban; calendar = month grid; absent means the default grid), the saved view they have applied, every saved view on the table, the active filter rules in plain English, and the columns hidden from them. rows is ALWAYS the whole table — filters are a display rule, not a read scope — so when the question is about what THEY see ("how many are left in my This Week view"), apply those rules to the rows yourself and say which view you counted. Change or save a view with eden_update_table. RECURRING ROWS: a row may carry repeats — a map of DATE property id → {freq, interval, from} (freq: daily | weekdays | weekly | monthly | yearly). Such a row REPEATS: marking it done rolls that date to its next occurrence and leaves the row open, so one row serves the task forever. Never add a row per occurrence to represent a recurring task, and never read a repeating row's date as a one-off. Set or clear a rule with the repeat field on eden_add_table_rows / eden_update_table_rows. repeats is absent on rows that do not repeat.
eden_resolve_creator
ChatGPTResolve a free-text creator query (handle, display name, profile URL) to one or more candidate social profiles. Use this when you're unsure which creator the user means and want to disambiguate before running an expensive analysis. Each result includes platform and username you can pass into eden_analyze_creator (via creatorRef) or eden_search_social_content (scope=creator).
eden_save_posts_to_board
ChatGPTSave indexed social posts onto a board in the user's Eden workspace as fully-hydrated cards (thumbnail, metrics, creator attribution — the same card Eden Discover saves). This is the save path for results from eden_search_social_content, eden_analyze_creator, and eden_analyze_list: pass each result's contentId (the Eden DB UUID, NOT the platform's native post id) plus its platform, verbatim. Typical flow: the user asks for outliers / topic posts -> a social search tool returns ranked results -> the user says "save those to my <X> board" -> resolve the board id (eden_search_workspace_items) -> call this once with all the posts. Posts land in the requested order in the board's first grid slots. The response carries saved + itemIds for posts that landed and notFound for requested ids Eden couldn't resolve (wrong platform, stale id). Retrying an identical call is safe — each post lands on a given board once, keyed by its contentId. Pass section to land the posts inside one of the board's existing sections.
eden_save_posts_to_library
ChatGPTSave indexed social posts to the user's Eden Library (their saved items, no board) as fully-hydrated cards, the same as Discover's "Add to library". Use this when the user says "save this" without naming a board; use eden_save_posts_to_board when they name one. Pass each result's contentId (the Eden DB UUID from eden_search_social_content, eden_analyze_creator, or eden_analyze_list) plus its platform, verbatim. The response carries saved, itemIds, notFound for ids Eden couldn't resolve, and libraryUrl to open the Library. Retrying an identical call is safe: each post is saved once per workspace.
eden_schedule_post
ChatGPTSchedule a social post in Eden and enqueue it for publishing — or, with draft: true, save it as an unscheduled scheduler draft for the user to review (no publish time, nothing enqueued). Scheduling is a real write, not a proposal card: schedule at a time only when the user explicitly asks to schedule/queue. When scheduling, pass a concrete timestamp as scheduledFor epoch milliseconds or scheduledAtIso — never natural language; when draft: true, omit the timestamp. Composing: preserve the user's line breaks exactly in text; a media-only post may use an empty text. Text-only works on X (twitter), Threads, LinkedIn, and Substack notes. Instagram and TikTok need hosted media (TikTok also takes a photo or a photo carousel, up to 35 images). media takes already-hosted public URLs from the scheduling media upload tools, never local paths. Shared images go to EVERY selected platform, LinkedIn included; never strip them from one unless the user asked. A PDF on linkedin publishes as a document carousel. Media for SOME platforms only: mark the asset scheduleMediaRole: "platformScoped" AND list its id in each intended perPlatform.<p>.mediaIds; perPlatform.<p>.mediaIds: [] makes one platform text-only. YouTube is Shorts only: one vertical video plus perPlatform.youtube.title. X / Threads threads: pass segments (too-long text without them is auto-split). X: text that is ONLY another post's x.com URL publishes as a native repost (needs X advanced automations; the response warns when they're off); commentary plus a post URL is a quote post. Long-form: article with explicit platforms from twitter / substack. Before targeting Substack, ask for audience, whether to send email, and whether to add a paywall; never infer them.
eden_search_creators
ChatGPTDiscover PEOPLE rather than posts from Eden's pooled creator embeddings. Use topic for "find competitors in this niche" / "who talks about X"; similar-to-creators for "creators like @name" using exact platform+username refs; or similar-to-list for more creators like one curated list. Results include profile data plus a reason or matching-post evidence. Use eden_resolve_creator only to identify a person already named, eden_analyze_creator to study one creator, and eden_search_social_content when the requested result is posts rather than people. Always send kind plus its required companion field: query, creatorRefs, or listId.
eden_search_instagram_music
ChatGPTSearch Instagram's licensed music library for a song to put on a single-video Reel. Returns tracks with audioId, title, artist, durationMs and coverUrl. Pass the picked track as perPlatform.instagram.music ({ audioId, title, artist }) on eden_schedule_post, eden_publish_post_now or eden_update_scheduled_post. Music needs a query (there is no trending list) and has no audio preview. Instagram may not offer every song in every country. Status needs_facebook means Instagram is connected with Instagram login: ask the user to switch it to Facebook login in Eden (Settings -> Social accounts, or the Music card in the post composer), then search again.
eden_search_social_content
ChatGPTSearch social posts across one of four scopes: a single creator, a curated list, every creator the user follows, or the entire indexed corpus. Optional free-text query enables semantic search; omitting it returns the top posts in the chosen scope ranked by orderBy. Use this to answer "what are the top posts from @X this week", "what's performing in my Y list", "find me posts about morning routines", etc. Each result carries a contentId for follow-up eden_read_social_post calls. Pattern-spotting filters (creatorTier, minFollowerCount / maxFollowerCount, minOutlierScore) apply to list / following / global scopes and are ignored on creator scope. For "what's working" / find-me-diamonds hunts default to creatorTier: "mid" + minOutlierScore: 2 (see server instructions). The response echoes appliedFilters. Empty topical results remain empty; follow the server research recovery guide. Relevance/deep supports topic, platform, dates, not Discover-only filters. Coverage describes a limited candidate page, not every post on a topic. Each call shows the user a row of post cards. Call it ONCE per question. Search again only if it returned no posts or the wrong topic. Results default to the user's language; pass languages only if they ask for others.
eden_search_workspace_items
ChatGPTText-substring search across the user's workspace items (notes, cards, boards, media, links). Matches case-insensitive against the item's title AND its url (when present) so 'stripe' finds both a note titled 'Stripe API' and a saved tweet from stripe.com. Optional type filter narrows by item kind. USE WHEN the user names a topic, project, person, or rough title and you want to find what they've already saved about it. Examples: 'find the onboarding note', 'where's my Stripe API note', 'do I have anything saved about morning routines', 'find that Loom about pricing'. Also answers 'have I saved anything by <creator>?' — search the creator's handle and/or profile domain as q (URLs are matched). For a RANKED analysis of a creator's full output use eden_analyze_creator instead — search returns saved cards, not ranked posts. NOTES: - Substring match only — no semantic search. The agent should pick distinctive query tokens (proper nouns, domain words) rather than common phrases. - Note BODIES are not searched (only titles + URLs). Once you've narrowed to a few candidate notes, follow up with eden_get_note_markdown to confirm the body matches. - Sidebar FOLDERS matching q come back in a separate folders field (name + the boards each holds). Folders are visual organization for BOARDS only and can never hold items — when the user says "in my <folder> folder", use a board from that folder's boards list or create one inside it with eden_create_board's folder param.
eden_study_top_carousels
ChatGPTResearch top-performing Instagram carousel posts and return a slide-by-slide teardown (structure, hook, per-slide text, design patterns) as reusable pattern notes. Study ONE creator's carousels (pass creator) OR a niche across creators (pass niche, optionally scope: "following" to stay inside the user's follows) — pass one or the other, not both. Billed research; vision analysis can take ~30s+. This does NOT generate carousel media — Eden exposes no AI media generation over MCP (image and carousel creation happen inside the Eden app).
eden_study_top_titles
ChatGPTExtract reusable TITLE SKELETONS from real top-performing headlines. Use this for "give me titles based on what is working" and YouTube title and packaging work. Returns, per winner: the syntactic FRAME with slots (After [studying N of a thing], here's how to [outcome]), the verbatim source headline, what each slot needs, the move that makes it pull, and doNotReuse — the source's own numbers, names, and story. HOW TO USE THE RESULT: reuse the frame, replace the substance. Do not read a skeleton, restate the principle behind it ("uses credibility and specificity"), and then write a title from that principle — that discards the structure and produces vague titles. Never carry a doNotReuse value into a title you write, and never present a skeleton as a finished title. ONLY THREE SOURCES have authored titles: YouTube videos, Substack ARTICLES, and native X ARTICLES. SHORT-FORM HAS NO TITLE — tweets, Substack notes, reels, TikToks, LinkedIn and Threads posts are excluded, because their first line is body text that happens to sit on top, not packaging someone wrote. The HOOK INSIDE a video (the spoken or on-screen opening) is also a different artifact and lives in the transcript. Study ONE creator (pass creator) OR a niche across creators (pass niche, optionally scope: "following") — one or the other, not both. Pass topic and each frame comes back already refilled as an example. Billed research.
eden_update_plan
ChatGPTAdd cards to the user's plan (their Planner) or change existing ones. Only on the user's ask ("add these to my plan", "move it to Thursday", "mark it posted"). Ideas the user picks to keep go on a board, not here: one eden_create_note per idea with presentation: "card" and the board as destination (ask which board when they haven't said). action: "add" takes items (1-20): text (the idea, one or two lines), and optionally stage (a stage name; default the first, its new-idea stage), date (YYYY-MM-DD), pillar (id or name from eden_get_strategy), format, platform, notes, research / draft (a doc or board id to attach), and the topic / angle it came from. Pieces (one level): plan long-form with the short pieces cut from it by giving the long card pieces (cards made under it; they take its pillar, topic, and angle unless set), or give a new card parent (an existing main card id), not both; 20 cards per call at most, pieces included. action: "update" takes updates (1-20), each with a card id from eden_get_plan and any of text, stage, date, pillar, format, platform (null clears a date, pillar, format, or platform), research (a board or doc id for its Research) and draft (a doc id for its Draft; null clears either), parent (a main card id makes it a piece; null takes it out), archived (true puts it off the board, false restores it), addNote (appended to the card's notes; notes are never replaced), or editNote ({ find, replace }: swaps one exact piece of the notes, to fill a blank or fix a line; eden_get_plan shows only the start of long notes, and a miss returns the full notes to copy from). Dates are the user's local days. To put a card on a pillar's day, read weekdays from eden_get_plan (which pillar runs which weekday) and pick the next open matching day. Taking a card off a planner with archived: true is the same as Remove from Table in the app, and it can be restored. This tool never permanently deletes a card. Returns the saved cards plus errors for any item skipped (by index); fix and resend only those. Planners: when the result has planner, the user's plan is a table in their Library (their plan tables are in planners: planners and any table with a Posts column, e.g. one per platform). Every table has its own columns and stage names (it can be anything), so go by what the result says, never by a template's names. Reads and writes default to the FIRST plan table; pass planner (a table's name or id) when the user names another table, and ask which one when they have several and it isn't clear. With no plan table, their old plan board moves onto a "Content planner" table on its own (once; old cards stay stored), or adding cards makes a fresh one (the note says which; pass it on). On a planner, stage is one of its stage column's options (stages, each with a role when it has one: new is where a new idea lands, review, scheduled, done; the others are working stages), date is its Date column, platform is its Platform column's choice, and parent isn't supported (pieces become their own rows with a note). archived: true takes a card off the planner, the same as Remove from Table in the app, only on the user's ask; archived: false brings it back. Those cards are left out of reads and counted in archivedCount, and reading with archived set to only lists them so you can restore one. An idea already on the planner (same words) is skipped, not added twice: the result lists it in skipped. A card's posts are its Publish posts (id, status, scheduledFor, platforms, preview). To post a planned idea, schedule or draft it with planCardId (the card id): the post joins the card's Posts column and the card moves itself (to whenScheduled, then to its done stage once it's out), so never move a card with posts by hand. A card's research (a board or doc) and draft (a doc) are the work it's made from: set them with the id you got making the board or doc (null clears), and open draft before writing a new doc for a card that has one.…
eden_update_scheduled_post
ChatGPTEdit an existing draft or scheduled post in Eden by id — change its time, its body, its auto first-comment, and/or its auto-repost. Find the id with eden_list_scheduled_posts. (1) reschedule only — pass scheduledFor epoch milliseconds or scheduledAtIso, leave the body fields out, and the existing content/platforms are kept; (2) edit the body — pass any of text, article, segments, media, platforms, or perPlatform and the post is rebuilt (omitted text/platforms fall back to the existing post, and existing media is kept unless you pass new media); (3) firstComment sets or clears the follow-up comment posted after publish; (4) autoRepost (X only) sets the auto-retweet cadence. Any combination in one call. The composing rules on eden_schedule_post apply, including the Substack delivery settings. Preserve line breaks exactly in text. Cannot edit a post that is already publishing or posted (returns a conflict). Never pass natural language for the time.
eden_update_table
ChatGPTChange a table itself (not its rows): rename it, ADD columns, or change the shared view (layout table/list/board/calendar — 'list' is the to-do shape — groupBy a column name or 'done', hide toggles). Existing columns can't be renamed or deleted here. Read the table first with eden_read_table; anything unresolvable comes back in warnings. SAVED VIEWS: view.saveAsView names the current setup so the user can return to it ("make me a This Week view" → set view.filters + view.groupBy, then saveAsView: "This Week"); view.applyView switches to one by name (null = All items); view.filters and view.hiddenColumns change what the table shows now, by column NAME. All of it is SHARED state — the whole workspace sees the table the way you leave it, so change the view when the user asks about their view, never to make your own reading easier. FILTERS HIDE ROWS, READS DON'T: eden_read_table returns EVERY row whatever the view filters. When the question is about what the USER sees, apply the rules yourself over the rows you read — the read's viewSummary names the active view and its filters.
eden_update_table_rows
ChatGPTUpdate rows in a table: set cells, rename, check/uncheck done, set or clear a repeating date, or soft-remove. Address a row by its row item id (from eden_read_table), its exact title, or its rev stamp from a CURRENT read (the stamp changes on every edit, so a stale one matches nothing — re-read then) — ambiguous selectors are skipped with a warning. Cell patches merge; null clears a cell. Pass each row's rev from eden_read_table as expectedRev: a row someone edited since that read is refused instead of overwritten and counted in conflicts — re-read the table and retry those rows.