Agent

Tool reference

Every tool the MCP server exposes, what it takes, and what it returns.

Tools are listed only when your token carries the scope they need. An agent that cannot see a tool will not call it.

Every tool calls the same REST route the API uses. There is no second implementation, so a limit enforced over REST is enforced over MCP.

Scopes

NameTypeDescription
accounts:readreadList connected channels.
posts:readreadRead publishing history, replies and insights.
posts:writewritePublish, delete, reply, hide and upload.
offline_accessoauthKeep the agent connected after the browser closes.

accounts:read reads connected channels. posts:read reads publishing history. posts:write publishes, deletes, replies and uploads. offline_access keeps the connection alive after you close the browser.

Tools

NameTypeDescription
list_accountsaccounts:readConnected channels with their ids. Call it before publishing.
select_channelsaccounts:readPick which channels to post to, and show the result to the person before publishing.
list_platformsaccounts:readEvery channel, its limits, and the JSON Schema of what options.<channel> takes. pricing is set only on X: what each X call costs from X credits, in mills (thousandths of a US dollar).
list_postsposts:readRecent publishes made through uplika, scheduled and draft posts included. status, since, until and sort=scheduled answer "what is going out this week".
list_channel_postsposts:readWhat is on the channel right now, including posts written in the app. Each item carries a permalink.
publishposts:writecontent, accountIds, optional mediaIds, and options for per-channel settings (the shape is in GET /v1/platforms under options; every channel takes content for a different text on that channel alone). scheduledAt holds the post and sends it at that time on any channel; draft: true saves it without sending. YouTube needs options.youtube.title and exactly one video. Instagram needs media on every post; a single video becomes a reel, and options.instagram.audio puts music from search_instagram_audio on it (accounts connected via Facebook only). Facebook publishes to a Page. Bluesky counts graphemes and also caps UTF-8 bytes; set options.bluesky.langs. On Naver Blog the body places media itself with ![alt](media:<id>) and @video(media:<id>), each alone on its line. Ghost needs options.ghost.title and places photos the same way with ![alt](media:<id>). WordPress needs options.wordpress.title, takes the same Markdown body as Ghost and refuses raw HTML.
open_postposts:readOne post's text, reply thread and metrics in a single call. The right tool when someone hands you a link. Optional accountId picks one channel's copy of a post that went to more than one channel. On X a low X credit balance leaves metrics null instead of failing.
get_postposts:readOne publish and the status of each target. X targets carry billing: { operation, mills, state, balanceMills }.
retry_postposts:writeRetry only the targets that failed. force: true retries a naver_draft_unknown, naver_reply_unknown, wordpress_publish_unconfirmed, clip_unconfirmed, x_publish_unknown, tiktok_publish_unknown, published_as_imported_copy or publish_outcome_unknown target. If another post with the same content and the same files (a file uploaded again counts as the same file) is already on the same account, it answers duplicate_post and retries nothing; force: true sends it anyway.
update_postposts:writeChange a scheduled or draft post: content, mediaIds, accountIds, options, threadItems, scheduledAt (null turns it into a draft) or draft. Fields left out keep their value. A Naver Blog post still waiting for the browser extension takes scheduledAt and draft the same way. A scheduled or draft chain can be changed too: content edits the first item, threadItems replaces the whole chain. Posts that went out can be rewritten in place on channels that support it (list_platforms feature update): a published Naver Blog post keeps its URL when you pass content, mediaIds or options (a post going only to Naver takes its media list from the new body's media: references when you leave mediaIds out; more than Naver takes answers naver_media_over_limit), and YouTube posts and the text of Facebook text, link, photo and reels posts are edited in place. A posted Naver Clip is edited in place in Naver Clip Creators by the browser extension (same video id, extension 0.7.46 or later): the description is rewritten from content with the post's tags kept unless you pass options.naver_clip.tags, and categoryId, visibility and disableComments change only when you pass them; changing the video, draftOnly, the cover or the content link answers naver_clip_field_not_editable. A clip still in drafts is edited the same way and stays a draft with the same id (extension 0.7.47 or later). A post written directly on the channel (origin imported) is read from the channel first, so fields you leave out keep the channel's values; it can be edited on YouTube, on Facebook text, link and photo posts, and on Naver Blog (feature update_imported), elsewhere post_not_editable. On Naver the post is rewritten from its current source, where the blocks already in the post (photos, videos, cards, tables, styled headings and quotes, whole text blocks with colors, sizes or alignment) are @keep(...) lines kept exactly as they are; new blocks, a new table included, go anywhere in the usual grammar, and to change an existing table you delete its line and write the table again in GFM; the first call answers naver_source_required with that source and its sourceVersion, and you send the edited source back as content with sourceVersion. That needs extension 0.7.8 or later. If the Naver edit screen shows a missing-image alert on an old post, this rewrites it in place and fixes it (same URL). Naver may answer 429 naver_rate_limited with retryAfterSeconds when its per-ID cap is hit. Editing a post that is already live is asynchronous: pass wait: true to hold the response for the result (up to 75 seconds), or read it with get_post. If the account that published the post was disconnected, this answers 409 account_disconnected; reconnect the same account and the post id comes back. Editing a live post rewrites every channel it went out on, so a link or a channel's own id on a post that is also on other channels answers 409 post_spans_channels with postId and channels and rewrites nothing; the same call with that postId rewrites every channel. While another update of the same post is still in progress, a published post answers 409 edit_in_flight and nothing is saved.
publish_nowposts:writeSend a scheduled or draft post right now. Same function the scheduler runs, so it cannot go out twice.
delete_postposts:writeDelete the post from every channel it went out on. A link or a channel's own id names one channel's copy, so when the post is also on other channels this answers 409 post_spans_channels with postId and channels and deletes nothing; the same call with that postId deletes it everywhere. On a scheduled or draft post nothing is on any channel yet, so this cancels it and removes our record. If a channel's trash is turned off, the post is gone for good and the response's note says so. The browser extension removes a Naver draft after the jobs ahead of it in that browser: if that has not finished within 60 seconds, the response's note says so and the extension finishes it within 30 minutes. If the account that published the post was disconnected, this answers 409 account_disconnected and deletes nothing on any channel; reconnect the same account first. If a delete of the same Naver post is still running, it answers 409 delete_in_flight and sends nothing new; for a draft, check list_naver_drafts for a draft with the post's title instead of get_post. Each X delete is charged the delete price from X credits; a low balance answers 402 x_credits_insufficient with topUpUrl before any channel is called.
list_repliesposts:readThe reply thread under a post. On a post that went to more than one channel, accountId and platform say which copy was read and otherAccounts lists the rest; optional accountId picks one. Naver Blog reads are done one at a time by the extension in the person's browser, so read Naver replies one post after another: send the next list_replies after the previous one has returned.
replyposts:writeid, content, optional replyTo for a nested reply, optional secret (Naver Blog), optional wait to hold until the reply is out (up to 10 seconds), optional accountId for which channel's copy on a post that went to more than one channel. replyTo also takes a triggerRef from list_automation_runs. On Naver Blog the id can be a link to someone else's post.
likeposts:writeid, optional replyTo to like a reply instead of the post, optional accountId for which channel's copy. Idempotent, no unlike. Works on Naver Blog, where someone else's post works by link, and on Instagram accounts connected through Facebook that hold the like permission. Any other Instagram account and every other channel answer not_supported. A Facebook-connected Instagram account connected without that permission answers reconsent_required: reconnecting it adds the permission.
followposts:writeblog, optional mutual; a mutual request needs message (up to 400 characters). Naver Blog only today: adds the blog as a neighbor, or sends a mutual-neighbor request that comes back pending.
find_naver_neighbor_candidatesposts:readquery, mutual, optional days (default 30) and accountId. Naver Blog only: blogs that recently posted on that topic, grouped by blog, with blogs this account is already connected to left out (mutual neighbors, requests it sent, refusals). Sends nothing. Each call uses one keyword measurement.
list_naver_neighbor_postsposts:readOptional neighbors ("mutual" by default, or "all") and accountId. Naver Blog only: neighbors' new posts from the last 48 hours, read from the account's neighbor feed through your uplika browser extension. Each post says whether its blog is a mutual neighbor (null when unknown) and whether Uplika already liked or commented on it through this connection. Sends nothing; like or reply to the posts you pick, one at a time.
hide_replyposts:writereplyId, postId, optional hide (default true), optional accountId for which channel's copy on a post that went to more than one channel.
delete_replyposts:writereplyId, postId, optional accountId for which channel's copy. Deletes the comment for good, which hide_reply does not. Instagram and Facebook remove anyone's comment on your post; on Threads and Bluesky a reply is itself a post, so only the connected account's own replies go. Read features in GET /v1/platforms for which channels have it at all.
get_insightsposts:readViews, likes, replies, reposts, quotes, shares. accountId and platform in the answer say which channel's copy was measured; optional accountId picks one. On X, metrics are switched on per X connection by the person on the Connections page: while off this answers 200 with noteCode x_analytics_not_enabled and charges nothing; while on, X is read and charged at most once per post per UTC day; a low balance answers 402 x_credits_insufficient with topUpUrl.
get_quotaposts:readHow much of the 24 hour allowance is used. Live usage from the platform, not the static limits. With an X account in scope it also carries xCredits: { balanceMills, topUpUrl }.
diagnose_naver_blogposts:readNaver Blog only: reads any public blog's RSS feed (latest posts, at most 50) and says whether its titles are written for search: the share with a search intent word, how many start with a date or episode label, posts per month, categories, and the words repeated in titles as candidates to research. Carries the date range it read. Cached 24 hours; a miss spends one of 20 diagnoses a day. This reads titles only and finishes at once; whether the posts are actually cited by the AI briefing or shown in search is scan_naver_blog_visibility, then get_naver_visibility with the blog.
expand_naver_keywordsposts:readNaver Blog only: expands up to 10 seed keywords one level through Naver autocomplete, returning each word with the seed it came from. The person's uplika Chrome extension looks them up in the background when it is on, otherwise our server does. When the answer is naver_autocomplete_unavailable, pass your own keyword list to research_naver_keywords.
research_naver_keywordsposts:readNaver Blog only: measures up to 60 keywords (searches, documents, posts per month), judges each, groups them into sets of one main plus two to five subs with the same intent, and gives every set a ready-to-paste Korean prompt. Runs against a time budget: partial: true with unmeasured[] means call again with the same keywords to finish from cache. Saved as a report (reportId), deduplicated within ten minutes.
list_naver_keyword_reportsposts:readThe person's saved keyword research, newest first, with each set and its prompt.
get_naver_keyword_historyposts:readOne keyword's measurement history, one point per measured day, newest first.
search_youtube_videosposts:readSearches YouTube for a keyword and returns the top videos with views, subscribers, length and publish date as YouTube gives them, the views-to-subscribers ratio and Shorts or long-form (both calculated by uplika and not YouTube metrics, which dataNotes states), plus a prompt in the person's Uplika language (Korean or English) that turns the table into title and hook ideas (pass topic to fill it in). The same search is cached 24 hours and does not count; a new one spends one of the person's daily searches and shared YouTube Data API units. Saved to the person's search history.
research_google_keywordsposts:readGoogle search volume for keywords from Google's keyword planner, for a country the person chose (country has no default: ask them first). Returns, per keyword, average monthly searches, the last 12 months (months once at the top, then one array per row), ad competition (competition LOW, MEDIUM or HIGH, not how hard the keyword is to rank for) and top-of-page ad bids in bidCurrency (the currency of Uplika's Google Ads account, the same for every country and not converted), plus related keywords with the same numbers, biggest first. A keyword Google has no data for is listed in noData, which is not the same as zero; a measured 0 means almost no searches. A field Google did not give is left out. Google rounds search counts into buckets, so they are approximate, and it splits some Korean words with spaces (the text is Google's and is not corrected). Put every keyword in one call: up to 10 keywords per search, as a list or one comma-separated string. Never call once per keyword; searches reach Google one at a time through a shared queue. The same search within 7 days is cached (cached: true). There is no per-person daily limit. Ask the person which keywords and which country before calling. While the tool is in beta the person can turn it off for connectors and API keys (research_tool_disabled).
research_adsposts:readPublic ads from Meta's Ad Library (Facebook, Instagram, Threads, Messenger), not the person's own ad account: what other brands are running, found by keyword (q) or by advertiser (advertiserId: the advertiser.pageId of a result ad, or a pageId from research_ad_advertisers). To see more ads by the advertiser of an ad you already have, pass that ad's advertiser.pageId as advertiserId. Returns up to 30 ads per page with the advertiser, title, text (cut at 500 characters), start date, runningDays, longRunning from 90 days, platforms, format, version count, call-to-action domain, thumbnailUrl (expires in a few days) and libraryUrl, plus received and kept counts when koreanOnly filtered ads without Korean text, a cursor for the next page, and today's usage. Meta publishes spend and impressions only in the EU, so those never appear; how long an ad has been running is the signal. Filters: country (default KR), status (active, inactive, all), media (all, image, video), sort (impressions, relevance; there is no newest-first sort, narrow with from and to dates instead), exact, koreanOnly. The same search within 24 hours is cached and free; each page that reaches the provider spends one of the person's daily ad searches. Show libraryUrl when quoting an ad and do not re-upload its images.
research_ad_advertisersposts:readFinds Facebook pages that match an advertiser name (q) in Meta's Ad Library, ordered verified first then by likes, with pageId, category, likes, Instagram handle and followers. Several pages often share a name, so show the candidates and let the person pick before calling research_ads with the pageId. Spends one daily ad search; the same name within 24 hours is cached and free.
research_ad_transcriptposts:readTranscript of a video ad from Meta's Ad Library by adId (from research_ads, hasVideo true). Answers available: false when the ad has no video, the video is longer than two minutes, no speech was found or the data provider could not read the video; asking again for the same ad gives the same answer. Saved for 30 days; a new transcript spends one daily ad search.
research_google_adsposts:readPublic ads from Google's Ads Transparency Center (Search, YouTube, Shopping, Maps, Play) for one website (domain) or one advertiser (advertiserId, AR followed by digits), not the person's own Google Ads account. The list has no ad text: each ad has the advertiser, format (text, image or video), first and last day shown, runningDays, longRunning from 90 days, and adUrl, the original in the Transparency Center where the person sees the ad without signing in. A domain result mixes resellers, partner brands and agencies under their own names (advertisersOnPage); pass an ad's advertiser.id as advertiserId to see only that company. Filters: region (default KR), platform (all, search, youtube, shopping, maps, play), format (all, text, image, video), from and to (ads shown at any time in that period). Up to 40 ads per page, most recently shown first, with a cursor. Google publishes no spend or impressions. The same lookup within 24 hours is cached and free; every other page spends one daily ad-analysis lookup, shared with research_ads. Show adUrl when quoting an ad and do not re-upload imageUrl images.
research_google_ad_advertisersposts:readFinds advertisers in Google's Ads Transparency Center by brand or company name, largest adsEstimate first, plus matching websites in domains. One company often has several advertisers, so show the candidates and let the person pick before calling research_google_ads with the advertiserId. A brand name that differs from the company name can find nothing; then ask for the legal name or the website. Spends one daily ad-analysis lookup; the same name within 24 hours is cached and free.
check_naver_visibilityposts:readLooks up to 10 keywords on Naver's integrated search through the person's uplika Chrome extension (it reads the public result page in their own browser, about 3 seconds apart, without cookies or a window) and returns, per keyword, whether an AI briefing answers it and in which form, its body and image sources with ranks and how many answer paragraphs cite each source, where blog posts sit in the integrated search results, and the related questions. Pass blog to mark that blog's posts and get its verdict. A keyword read in the last 30 minutes comes back from the saved result unless fresh is true; new reads spend the person's daily allowance of 200 Naver search pages. It waits up to 60 seconds and answers partial with pending when the rest is still being read in the background. The answer also carries citationRules: sentence rules uplika measured on posts the AI briefing quoted, and weaker suggestions (grade B) to offer rather than check, each with its evidence and caveat (not Naver's criteria). Each keyword result also carries organicRead: when it is false, the integrated search blog and web sections could not be read for that keyword (the page shape changed or the result was cut), so exposed: false and an empty organic list mean nothing was read there, not that the post is absent; say so instead of reporting no exposure. With blogTab: true it also reads, for each keyword, page 1 of the blog tab sorted by relevance, and adds blogTab (read, size, the top 5 posts with mine marked, reason) and verdict.blogTabRank: the blog's place among the blog posts on that page (ad entries in the list are not counted). That is one more search page per keyword, so it spends twice the daily allowance and takes about twice as long; up to 7 keywords with blogTab finish within the wait. blogTab.read: true with blogTabRank: null means the blog is not on page 1 (size is how many posts that page had). blogTab.read: false means the blog tab was not read and reason says why; reason extension_outdated means the person's extension is older than requiredVersion, so only the integrated search was read. Results for local keywords can differ with where the person's browser reads from.
scan_naver_blog_visibilityposts:readStarts a report card for a Naver blog: reads its latest posts (up to 50, the default) from the public RSS feed, picks each post's main keyword and checks them through the person's uplika Chrome extension in the background. Unlike diagnose_naver_blog, which reads titles only, this searches each post's keyword; it also refreshes the title summary without spending a diagnosis. Answers runId and etaSeconds at once; a report card for the same blog from the last 6 hours is returned instead (reused).
get_naver_visibilityposts:readReads what uplika recorded about Naver search and AI briefings without reading Naver again: with blog, the report card (four number cards, five scores on uplika's own scale with raw values and thresholds, the lowest axis, each post's latest verdict, events); with post (a Naver post address or an Uplika post id), that post's check history and events; with runId, one check's results so far. With blog, the cards count every post of that blog uplika has checked so far, not only the last scan; lastScan.total is the size of the last scan. Of the five axes, search_title, consistency and topic_focus come from the post titles; ai_cited and search_exposed come from search checks. Each keyword result also carries organicRead: when it is false, the integrated search blog and web sections could not be read for that keyword (the page shape changed or the result was cut), so exposed: false and an empty organic list mean nothing was read there, not that the post is absent; say so instead of reporting no exposure. The report card's exposure rate and search_exposed axis count only posts whose sections were read (organicReadPosts); a post or trend day with organicRead false is unread, not unexposed.
get_publish_optionsposts:readWhat a channel needs to know before you publish to it. TikTok only: the creator nickname the post goes out as, the privacy levels this account may use right now, whether it can post at all, and its video length limit. Other channels answer not_supported.
search_instagram_audioposts:readInstagram only. Music or original sounds one connected Instagram account can put on a reel, for options.instagram.audio.id. Without query it lists what is trending for that account (it differs per account); with query, music comes only from Meta's royalty-free Sound Collection, so an artist or song name will not find that song. Instagram gives no sort or date option, no count of reels using a sound and no creation date; keyword search is strict about spelling and spacing, and licensed songs in the trending list come with listenUrl only, with no audio file. Works only for accounts connected via Facebook; others answer not_supported. Pass nextCursor back as cursor for the next page.
get_helpaccounts:readTroubleshooting answers: why something did not work and what to do. Pass code (an error code), query (a few words in any language), id, platform or locale; without arguments it lists every question. The same answers are on the Troubleshooting and FAQ page, and errors that have one carry its address in error.help.
report_problemaccounts:readSends a problem report to the Uplika team when get_help did not solve it and the person wants the team to look. personSaid is the person's own words and summary is what you tried; uplika attaches its own record of their recent failed requests and the reply goes to the email on their account. Show the person what you will send first. A repeat of the same problem within a day is recorded but not mailed again.
describe_grammarposts:readHow to write the body for a channel that has its own markup. Naver Blog has one; every other platform answers not_supported. Without a topic you get an overview and the list of topics (directives, attributes, inline, blocks, highlight, media, limits); with one you get that section in full, including the mistakes that fail silently. The values come from the same grammar the publisher validates against.
naver_layoutposts:readNaver Blog only: previews how publish will lay the body out (the house form: underlined quote headings, a rule between sections, one photo per section). Text never changes. Pass options.naver_blog.layout: "as-is" to publish exactly what you wrote. Pass forms: "all" to get every preset laid out side by side and let the person choose.
bridge_statusposts:readNaver Blog and Naver Clip only: whether the browser extension is online, when it was last seen, and per-OS commands that open Chrome in the profile that has the extension. publish and 503 bridge_offline carry the same object.
list_naver_draftsposts:readNaver Blog only: the drafts (temp-saved posts) in the blog's draft box with logNo, title and last-saved time, newest first. A post published with options.naver_blog.draftOnly is one of them, and so is anything saved by hand in the Naver editor. Needs extension 0.7.2 or later; older ones answer 422 extension_outdated.
publish_naver_draftposts:writeNaver Blog only: publishes a draft from the draft box exactly as it is in Naver, keeping edits made by hand in the editor. Do not send content; category, tags and openType come from the draft unless you pass them. Same result shape as publish. A draft made through uplika (externalId draft:<logNo>) flips that post to published instead of creating a second one.
refresh_accountposts:writeRe-read a channel's metadata. On Naver Blog this re-reads the categories through the browser extension and waits up to a minute for it. On Threads, Instagram and Facebook it re-reads the display name and the profile picture; it does not renew a sign-in, and a connection whose sign-in the channel has ended answers account_expired.
create_connect_inviteposts:writeCreate a connection link for someone without an Uplika account, such as a client, to connect their own channel account to this workspace from their own device. The link is only in this answer. Which channels a link can connect today is in list_platforms under connectInvite; others answer invitation_channel, and channels whose posts are written by the browser extension answer invitation_bridge. Listing and cancelling links happen in the dashboard.
media_upload_linkposts:writeAsk the person to upload from their own device. Returns a short-lived link to hand them.
media_upload_statusposts:readHas the person uploaded yet? Returns waiting, ready or expired, plus every media id uploaded through that link. Ready images also come back as image blocks (previewIds keeps the order) so the agent can see each photo; a media with duplicateOf is the same bytes as another id
media_from_urlposts:writeAttach a file that is already on the public web. One step, no upload. Share links from Drive and Dropbox return HTML, not the file.
media_from_fileposts:writeChatGPT only. Takes files from the conversation and stores the bytes unchanged: files the person attached (also from their ChatGPT library), and images ChatGPT made or files made by code when ChatGPT passes their /mnt/data path. A file id alone is refused.
media_presignposts:writeGet a media id and a one-time upload URL.
media_completeposts:writeConfirm the upload before publishing.
update_mediaposts:writeid, aiGenerated. Marks or clears the AI-made flag on an upload. Posts published or edited afterwards follow it: Naver Blog photos get the AI usage label, Instagram, YouTube and TikTok get their AI declaration. Posts already out do not change.
fit_ad_imagesposts:writeitems (mediaId, slots, mode, focus, padColor). Copies of images in the shape each Google Ads image field needs: landscapeImages 1.91:1, squareImages 1:1, portraitImages 4:5, tallPortraitImages 9:16 (Demand Gen only), logos 1:1, landscapeLogos 4:1. Photos are cropped and logos padded unless you choose; nothing is enlarged. The originals stay, each copy is a new media id with a url to look at first, and an image that already fits comes back unchanged. No ad account connection is needed.
list_automation_templatesposts:readEvery automation template with the params it takes. Start here before create_automation.
create_comment_to_dmposts:writeThe common one: someone comments, the account DMs them. Creates a draft; nothing goes out until enable_automation. Instagram can gate delivery on following; Facebook cannot and Threads has no DMs.
create_automationposts:writeCreate a draft from any template in list_automation_templates. For a flow no template covers, build the document and call put_automation on the draft.
list_automationsposts:readAutomations in the workspace with their status, trigger kinds and run counts.
get_automationposts:readOne automation as a document: triggers, nodes, start, plus the version put_automation needs.
put_automationposts:writeEdit an automation by replacing its whole document, for flows no template covers or already edited on the canvas. A flow that still has its template form answers automation_is_template: change it with update_automation, or pass detachTemplate: true. Send the version you read or it answers version_conflict.
enable_automationposts:writeTurn it live. This is the moment messages start going to real people, so confirm first. Answers reconsent_required when the account lacks a permission the flow needs, the DM permissions or the own permission of a mention trigger or of liking a comment on Instagram (reconnect it), and automation_catch_all_taken when another message flow without keywords is already live on the account.
disable_automationposts:writeStop it. Runs already waiting for a button stay waiting; nothing new starts.
send_to_past_commentsposts:writeRun a live comment automation on comments already on the post, the ones it missed. mode preview counts who would get it and who would not (older than 7 days, already answered, keyword, the same person's other comments), start queues the sends a few at a time, stop halts them. Instagram and Facebook. Preview first and confirm with the person before start.
update_automationposts:writeEdit an automation that still has its template shape: pass only the template params you want to change (objects such as deliver merge one level deep, null removes a field), plus the version you read. Also renames and turns it on or off. A flow edited on the canvas answers automation_not_template; use put_automation there.
delete_automationposts:writeDelete an automation and its runs. A live one is refused with automation_live unless force is true.
duplicate_automationposts:writeCopy an automation as a new draft with the same settings, optionally renamed.
validate_automationposts:readCheck a flow document for an account without saving it. Answers the same problems put_automation would.
list_automation_versionsposts:readSaved versions of one automation; read one with get_automation and version.
list_automation_runsposts:readRuns of one automation with why each one stopped: the window, an opt-out, a human reply, the channel, or the AI step.
list_conversationsposts:readThe inbox. Each conversation says whether the 24-hour messaging window is open.
read_conversationposts:readOne conversation with its messages, the contact, and how long the window has left.
send_dmposts:writeReply inside the open window. Refuses with window_closed outside it, with thread_owned_elsewhere when another app controls the conversation, with dm_access_disabled when the account owner turned off message access for connected tools in the Instagram app, and with dm_link_sharing_restricted when Instagram is not letting the account share links in messages (a message with a link in its text gets the same answer while that lasts); the account cannot start a conversation.
get_contactposts:readOne contact: name, tags, opt-out, window, and on Instagram whether they follow the account. Follower lists do not exist on any channel.
list_mentionsposts:readThreads only: posts where other people mentioned the account, from the mentions webhook. Mentions need a permission of their own; when no connected account in scope holds it, the list is empty and the response carries a notice with code reconsent_required, because reconnecting the account adds it.
approve_replyposts:writeThreads only: approve or ignore a reply held by reply approval. Omit replyId to read the queue.
{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "publish",
    "arguments": {
      "content": "Shipped nested replies today.",
      "accountIds": ["acc_01H..."],
      "mediaIds": ["med_01H..."]
    }
  }
}

Common questions

Why can my agent not see a tool?

Because the token does not carry that tool's scope. Tools are listed per scope, so a read-only connection never shows publish at all rather than failing when it is called.

Which tool should the agent call first?

list_accounts. publish takes account ids, not handles, so an agent that skips it either guesses or asks you a question it could have answered itself.

Do the tools behave differently from the REST API?

No. Every tool calls the same route REST does. There is no second implementation, so a limit enforced over REST is enforced over MCP.