How-To Guide
What can a comments and messaging API for social media automate?
What comment and direct message APIs on Instagram, Facebook, YouTube, X and TikTok can read, reply to and moderate, and what should stay with a person.
By the Superdirector editorial team. Platform documentation read and stored November 17, 2026. Vendor pages read the same day and dated in the text where the date matters.
At a glance
What can a comments and messaging API for social media automate?
A comments and messaging API automates reading and replying, not deciding. It can list the comments on a post, pull a conversation, fetch messages, reply inside the window the platform allows, and hide or delete a comment where the account owner has the right. The boundary is the same on every platform: the person must write first, an automated chat must say it is automated where the law requires it, and replies outside the window fail or are restricted. Complaints, money, refunds, legal and health claims, threats and anything that changes terms stay with a person. Treat the API as a reading and routing layer, and keep the judgement about what to say in human hands.
- Meta gates comment and message access behind permissions and, for other businesses' accounts, Advanced Access and App Review.
- The messaging window opens with the person: 24 hours for most actions, up to 7 days after an ad click, and 7 days for one private reply to a comment.
- YouTube charges quota units (1 for comments.list, 50 for comments.insert), while X charges per resource read and per request written.
- TikTok is the unresolved case: a 2023 vendor claim of comment calls against a 2026 guide that says no public endpoint exists.
Original diagram from platform documentation stored November 17, 2026. Blank cells mean the stored pages did not state the ability; the TikTok row carries the source conflict.
Make it work for your brand
Find it. Remake it. Post it. Superdirector turns the viral formats in your niche into your brand's carousels, X posts and video scripts, and keeps the calendar full.
The short answer: four surfaces, one boundary
A comments and messaging API does three jobs well and one job badly. It lists the comments on a post, pulls a conversation and fetches a single message. It replies inside the window the platform allows and only with the permission the platform requires. It hides or deletes a comment where the account owner has that right. What it does badly is decide whether a reply is yours to send, and that is the part most guides skip.
The boundary is set by the platforms, not by the vendor you buy. On Meta's messaging platforms a person must initiate the conversation, and the reply is limited to a window that opens with the person's action. The same rule appears in TikTok's business messaging documentation: a conversation must always be started by the user. An automated chat also has to say it is automated where the law requires it, and Meta's responsiveness policy gives automated bots 30 seconds to answer any input.
The permission gate is the second boundary. Meta separates comment access from message access, and an app that wants to use them on other businesses' accounts needs Advanced Access, which means App Review. Blotato's own announcement describes that work as the slow part: register a Meta app, justify each permission, wait out the review, then handle token refresh and scopes yourself. Builders who skip that step usually find the endpoints return nothing, not an error, which is worse.
So the honest framing is this. An API is a reading and routing layer. It can watch every comment and message, sort them, answer the repeated factual question, and put the hard cases in front of a person with the context attached. It is not a replacement for the person on the other side of a complaint.
What each platform's API lets you read and send
Coverage differs by endpoint, not by platform name. A vendor that says it supports Instagram has told you nothing about whether it returns comments, hides them, or can read a message thread. Here is what the platform pages actually document.
Instagram. The comment node is readable with the comment management permission, and it exposes the text, timestamp, likes, the parent comment for a reply, and a hidden flag. The limits matter more than the fields. Comments on age-gated media are not returned, comments on live video can be read only while the media is still being broadcast, and comments from restricted users do not come back unless those users are unrestricted and the comments are approved. Deletion is narrower than most builders expect: a comment can be deleted only by the owner of the media, even when the person asking wrote the comment, and hiding a comment does not hide the media owner's own comments. Read the Instagram comment reference before you promise a moderation feature.
Facebook Page. The comment object carries two fields that make comment-to-DM flows possible: can_reply_privately, which says whether the Page viewer can send a private reply, and private_reply_conversation, which returns the conversation a private reply opened. Hiding is a Page-only action. Rate limiting is the practical ceiling: the page lists error 80001 for too many calls to the Page account and tells the caller to wait, and error 283 for a missing page permission. See the Comment reference.
YouTube. Two endpoints cover most of the job. comments.list returns comments and costs 1 quota unit, with a page size of 1 to 100. comments.insert creates a reply and costs 50 quota units, and top-level comments go through a different method. Read the insertion limits carefully: replies to comments on posts are in a pilot with select partners, and broader public access is planned for general availability in 2027. YouTube also filters before your code sees anything, because its own moderation settings can hold comments in Studio for up to 60 days, hold comments that match blocked words, and hide a commenter across every video. An API-only automation reads a set the platform already filtered. See comments.list, comments.insert and YouTube's comment settings.
X. X exposes direct messages; the pages read here show no separate comment endpoint. The published DM contract lists nine operations, and sending goes through a conversation path with an OAuth2 user token that carries the dm.write scope. X prices this per unit and shows rate limits per endpoint per 15 minutes, with billing and rate limits described as separate controls. See the DM operations list, a current send guide and X's rate-limit page.
TikTok. This is the row where the sources disagree, and the disagreement is worth stating plainly. Ayrshare announced get, add and delete comment calls on TikTok posts in April 2023 through a partnership. A guide updated in May 2026 says TikTok has no public API endpoint for reading, hiding, deleting or replying to comments, and that the only API-level control is allow_comment at post creation. TikTok's own business comment page stored as a script shell, so no official text was readable. Messaging has a clearer path: Infobip documents TikTok business messaging with webhooks and replies, and MessageGate lists itself as an early access partner, both under the rule that the user starts the conversation. Read the Ayrshare announcement and the 2026 guide and test the account's own app before you build.
The rules that decide whether a reply is allowed
Reading is cheap and mostly unregulated. Sending is where the rules live, and they are the same handful of rules on every platform in this family. Learn them once and the implementation becomes a set of checks, not a set of guesses.
Meta's messaging policy sets the windows. The standard messaging window opens with a user action and, for most actions, lasts 24 hours. A first message after a click to a Messenger or Direct ad can open a window of up to 7 days. A private reply to a post comment is allowed within 7 days of the referenced action. Outside those windows, three mechanisms exist and each has a condition: message tags for approved use cases, a Human Agent tag that gives a separate 7-day period, and a one-time notification token that can be used once and expires within 1 year. Using tags outside their approved use cases may restrict the ability to send messages.
The platform overview adds the operational details that break naive implementations. A person must initiate the conversation. Automated bots must respond to user input within 30 seconds. High-volume messaging can temporarily disable inbox functionality. Where the law requires it, automated chats must disclose that the person is interacting with an automated service, and Meta names California and Germany as markets to pay attention to. Read the Messenger Platform overview alongside the policy.
One warning from a vendor is worth repeating because it changes the design. Blotato writes that sending outside an allowed window lands the message in failed status, and that it does not backfill history, since capturing starts when the account connects. If your agent assumes a send succeeded because the call returned, it will mislead the person watching the inbox. Check the result, not the request.
Original diagram of Meta's messaging windows and the ways back outside them, from the messaging policy and platform overview read November 17, 2026. Bars show documented duration, not scale.
What an agent can do without a person
The safe division of work is simple to state. The agent reads everything, sorts everything, and sends only what it can verify against stored facts. A person owns every message where the answer is a judgement, not a lookup. That division is also what the platforms allow, because a reply has to fit a window and a permission before anything about its wording matters.
Five jobs survive that test.
- Sort the inbox. Classify each comment or message as a product question, a complaint, a price or deal question, abuse or spam, or praise. The classification is what makes a human queue usable, and it needs no send permission.
- Answer a repeated factual question. Opening hours, where to buy, what a product contains, how a return works, and similar lookups can be answered from a stored source with a link. The answer must quote the source, not improvise around it.
- Route the hard cases. A complaint, a refund, a legal or health claim, a threat, and any conversation about terms go to a person with the thread attached, so the handoff does not restart the conversation.
- Moderate obvious spam. Hide or delete only where the platform and the account owner permit it. On Instagram that means the media owner deletes, not the commenter, and hiding does not touch the owner's own comments.
- Draft, do not send, outside the rules. When a reply falls outside the window or needs a permission the account lacks, the agent drafts it for a person instead of attempting a send that will fail.
The same split applies to how the account sounds. Meta requires automated chat disclosure in some markets, and the decision about whether an account presents itself as automated is a brand decision. An agent can insert the line a person approved. It should not decide that the line is unnecessary.
What must stay with a person
The list of human work is shorter than most teams expect, and it is not negotiable. It is where a wrong sentence costs money, trust or legal exposure, and where no stored source can settle the answer.
- Complaints and anything that might become one. An automated reply cannot verify what happened, promise a remedy, or apologise on the brand's behalf.
- Money, refunds and terms. Prices, discounts, contract changes and delivery promises are commitments. A classifier that confuses a question about price with a decision about price is the defect to prevent.
- Legal, health and safety claims. These need a person who can be accountable for the sentence, and often a review that is outside the inbox.
- Threats, abuse directed at staff, and anything that needs moderation judgement. Hiding a comment is a platform action; deciding whether to answer it in public is not.
- The disclosure decision. Whether an account says it is automated, and how, is a brand and legal decision that Meta's policy makes explicit in some markets.
- Any first contact the platform forbids. Where the person must write first, an automated outbound message is not a clever workaround. It is a failed send or a policy problem.
There is a quieter reason to keep these with a person. A reply that is fast but wrong is worse than a reply that is slow and right, because the wrong one is public and quotable. The agent's value is that it finds the hard cases quickly and hands them over with context, not that it answers everything.
Original diagram of the split this guide recommends, built from Meta's messaging policy rules stored on the same day.
What it costs, priced by the unit
Two platforms publish prices in units you can compare; Meta does not publish a per-call price on the pages read here. Use the table for the units that are documented, and treat a vendor's per-message quote as a different unit until you convert it.
| Platform | Operation | Documented unit | Price or cost |
|---|---|---|---|
| X | Read a post | Per resource returned | $0.005 |
| X | Read a direct message event | Per resource returned | $0.010 |
| X | Create a DM interaction | Per request | $0.015 |
| YouTube | comments.list | Quota unit | 1 unit per call |
| YouTube | comments.insert | Quota unit | 50 units per call |
| Meta | Comment and message calls | Not published on the pages read | Rate limits and App Review, not a fee |
The X numbers come from its pricing page, which states that reads are charged per resource fetched and writes per request. The YouTube numbers come from the list and insert pages, which give the quota cost at the top. Meta's cost is time and permissions, with no published fee, and its own pages warn that too many calls to a Page account produce error 80001.
The trap is unit mixing. A third-party DM tool advertises $0.010 or more per DM, while the official page charges $0.015 per request to create a DM interaction and $0.010 per DM event read. Those numbers are close enough that a comparison table can look consistent while measuring different things. Convert everything to the same unit before you recommend a route, and read what the X API actually costs when the plan depends on volume.
Volume also changes the shape of the bill. X caps pay-per-usage plans at 3 million post reads per monthly billing cycle and points higher volume to an Enterprise plan, and its rate limits apply per endpoint per 15 minutes on top of the bill. A design that polls dm_events every minute will hit the limit before it hits the budget. Read the rate-limit page before you pick a polling interval.
Where Superdirector draws the line
This guide is published by Superdirector, so the scope belongs in the open. Superdirector runs publishing and research for a brand's social accounts. It does not answer comments or direct messages, and buying it for that job would be a mistake.
Growth is $199 a month for one AI-operated account: scheduled research, routines and supported Instagram and X publishing, with 2,000 credits a month and extra accounts at $99 each. Starter is $29 a month for people who work from Claude or ChatGPT: live social data, ideas, scripts and downloads. The pricing page lists both, and the $199 figure buys one AI-run account, not a team seat.
The MCP connection has its own boundary, and it is the relevant one for this question. It reads the Canvas's saved sources and brand context and saves drafts, starts credit-using work only after a credit quote and the person's OK, and posting is approved in the Canvas, not from the chat. Autonomous channels, AI-generated video and video posting are not available through the connection. Those are the words in the MCP guide, and they are why an agent working from Superdirector does not reach a comment endpoint.
If you are choosing between an API you build and a product you connect, the question is whether you want to own permission reviews, token refresh and window logic. Read social media API versus MCP for founders and the MCP server roundup for that comparison. If you do build the comment layer, what the official APIs publish about rate limits is the next page to read.
A launch checklist you can run this week
This is the order that avoids the two expensive mistakes: building against an endpoint the account cannot use, and sending a reply the window forbids.
- Pick one surface with a documented endpoint. Instagram or Facebook comments, or YouTube comments, are the safest first builds. Do not start with the TikTok comment question until you have tested the account's own app.
- Apply for the permissions before you write the feature. Meta separates comment and message access and needs Advanced Access and App Review for other businesses' accounts, so the wait is the first thing on the plan.
- Store the page token and the webhook. Meta triggers a webhook when a message arrives, and polling is the fallback, not the design. Confirm the token refresh path before the first deploy.
- Classify before you answer. Build the five kinds (product question, complaint, price or deal, abuse or spam, praise) and store the label with the message. Everything downstream depends on it.
- Enforce the window in code. Refuse to attempt a send outside 24 hours, the 7-day ad window, or the 7-day private reply, and route those to a person instead.
- Log every send with a receipt. Record the request, the returned status, and the message id. A failed send outside a window is a business fact, not an exception to swallow.
- Test on one account for a week. Watch for the silent failure modes: an endpoint that returns an empty list because the permission is missing, and a send that returns success while landing in failed status.
- Publish the escalation route. Tell the person who receives escalations what lands in their queue, with the thread attached, and how fast the platform's 30-second expectation applies to the automated part.
None of this needs a new platform. It needs the willingness to keep the send button behind a person for the cases that matter, and to let the agent do the reading that makes those cases cheap to find.