Messaging
AI community manager that replies to social media comments in your brand voice
Community Manager reads new comments and mentions on your connected accounts in one batch, classifies sentiment and intent, answers questions from your knowledge base, and flags complaints, refund requests and legal risks for you with the reason.
- As a workflow node
agent.communityManager- Default model
- Claude Sonnet 5
- Economy model
- Claude Haiku 4.5
- Tools
- 5
- Report
- Chatbot insights
What the agent does
To auto-reply to comments, the agent works with the platforms that support reading and replying to comments: Instagram, Facebook, X and YouTube. LinkedIn, Threads, TikTok and Pinterest don't offer this. A saved agent remembers the time of the last comment it handled, so the next run only fetches newer comments and nothing is processed twice.
All new comments are classified in one batch for sentiment (positive, neutral, negative), intent, urgency and language. For product, price, shipping and policy questions the agent searches your knowledge base first and answers only with what it finds; if there is no answer it says the team will follow up instead of guessing. Replies are 1–2 sentences, in the commenter's language and specific to what they said; spam and emoji-only comments are skipped.
Comments that contain an escalation keyword, are negative with high urgency, report harm, carry legal or PR risk, or come from the press are flagged for you with a severity and a reason; the agent still prepares a short, empathetic holding reply for them. For a ready-made hourly flow, see the comment and mention auto-reply template.
Inputs / Outputs
Inputs
- Auto-reply without approval
- When off (default), the agent posts nothing; every reply is returned as a draft in the output and report for you to review. When on, replies are posted under the comments directly. Keep it off for the first 1–2 weeks, until you trust the quality.
- Escalation keywords
- Words that send a comment to you when they appear. The default list is
refund,lawyer,complaint, and the agent also considers their Turkish and English equivalents. Add your own terms and spelling variants, e.g.chargeback,scam,consumer protection.
Outputs
- replies
- Every comment that got a reply or a reply draft: platform, author, comment excerpt, reply text, sentiment, intent and whether the reply was posted.
- escalations
- Comments flagged for you: platform, author, reason and severity (low, medium, high).
- sentiment
- Sentiment totals across all new comments processed, not just the answered ones: positive, neutral and negative counts.
Saves a report to the Reports page on every run: Chatbot insights.
Tools
- Read comments
get_comments - Reply to comments
reply_comment - Sentiment analysis
classify_sentiment - Hand off to a human
escalate_to_human - Knowledge base search
knowledge_search
Integrations it uses
Example tasks
- “Reply to new comments in at most 2 sentences, warm but respectful. Use the knowledge base for shipping and return questions; if unsure, say the team will follow up.”
- “Send me the “discount code doesn't work” comments on the campaign post. Answer the other questions and remind people the campaign ends on 31 January.”
- “Don't reply to spam or competitor ads, just flag them. Thank people who praise us with a short note using their username.”
- “Don't get defensive on negative comments; apologise and ask them to contact us so we can fix it. Never haggle over price.”
Tips for good results
- Put tone and form of address in the project's Brand voice field (formal or casual, emojis or not); the agent sees it on every run.
- Add FAQ, shipping, returns and warranty documents to the knowledge base. If an answer isn't there, the agent won't make one up; it says the team will follow up.
- Extend the escalation list with spelling variants and local terms. If the list is too broad, every bit of criticism lands on your desk.
- Write legal limits in the saved agent's Additional instructions: no health claims, no exact delivery dates, never ask for personal data in a public reply.
- In an hourly flow, every run that finds comments writes a report; switch off Save report on the node and leave the weekly summary to Report Writer.
Model advice
Default model
Claude Sonnet 5
Per 1M tokens: $2 input · $10 output
Economy model
Claude Haiku 4.5
Per 1M tokens: $1 input · $5 output
The default balanced model is recommended for getting the tone right on negative or sarcastic comments and protecting your brand voice. If most comments are short, repetitive questions about shipping, prices or thank-yous, the economy model is enough. Sentiment classification always runs as a separate, low-cost economy-model call.
Templates that use this agent
Frequently asked questions
On which platforms can it reply to comments?
Instagram, Facebook, X and YouTube; Instagram needs a professional account. LinkedIn, Threads, TikTok and Pinterest don't offer comment reading and replies, and private messages such as Instagram DMs are outside this agent's scope.
Does it post replies without me seeing them?
Not by default: with auto-reply off, replies stay as drafts. Turn on Auto-reply without approval once you trust the quality; the ready-made template ships with it on.
Does it delete or hide negative comments?
No. The agent never deletes or hides comments; it flags sensitive ones for you with the reason, and you decide.
Does it answer WhatsApp messages too?
No. One-to-one messages are handled by AI Chatbot; Community Manager works with public comments and mentions.
Get started with OrqLabs
Build your first workflow today
Start on the free plan: connect your accounts, pick a template and stay in control with approval steps. No credit card required.