How to Let AI Agents Answer DMs and Comments via MCP
Give Claude or any MCP client the tools to read and answer Instagram DMs, moderate comments, and reply to Google reviews — engagement tools on the Postproxy MCP server.
Beyond publishing
The Postproxy MCP server is mostly known for letting agents publish posts. The same server exposes the full engagement surface — DMs, comments, and Google Business reviews — as tools any MCP client (Claude Code, Claude Desktop, Cursor, or your own agent loop) can discover and call.
Setup is the same one-liner; hosted, no install:
claude mcp add --transport http postproxy \ https://mcp.postproxy.dev/mcp?api_key=YOUR_POSTPROXY_API_KEYThe engagement tools
Direct messages
| Tool | What it does |
|---|---|
dm_chats_list | List conversations for a profile |
dm_messages_list | Read a thread |
dm_message_send | Send a reply |
dm_message_react / dm_message_unreact | React to a message (Facebook & Instagram) |
dm_comment_private_reply | DM a commenter (private reply) |
Comments
| Tool | What it does |
|---|---|
comments_list | List comments on a post |
comments_create | Reply (or comment on the post) |
comments_hide / comments_unhide | Moderate without deleting |
comments_like / comments_delete | Like (Facebook) or remove |
Google Business reviews
| Tool | What it does |
|---|---|
profile_comments_list | List reviews, filterable by location |
profile_comments_create | Reply to a review |
profiles_placements | Enumerate locations |
What a session looks like
With the server connected, engagement work becomes prompts:
“Check the Instagram inbox. Draft replies to every unanswered DM from the last 24 hours, show them to me, send the ones I approve.”
The agent calls dm_chats_list, reads threads with dm_messages_list, drafts, and sends with dm_message_send after your approval.
“Go through comments on yesterday’s launch post. Hide anything that’s spam, reply to genuine questions, and list what you did.”
That’s comments_list → judgment → comments_hide / comments_create — the agent version of the moderation pipeline, with the classifier and the actuator in one place.
“List this week’s Google reviews for the downtown location and reply to the 5-star ones.”
profiles_placements to find the location, profile_comments_list with its placement_id, then profile_comments_create per review.
Guardrails that actually work
Agents with send-capable tools need boundaries sharper than a system-prompt plea:
- Start read-only. Let the agent run on
*_listtools for a few days; review what it would have sent before granting the write tools. - Approval gates on send. MCP clients like Claude Code prompt per tool call by default — keep
dm_message_send,comments_create, andprofile_comments_createbehind manual approval until trust is earned. Hide/unhide is reversible and safe to allow earlier. - Platform rules still apply. Meta’s 24-hour DM window, one private reply per comment, one owner reply per Google review — the API enforces these, so a confused agent gets a clean error instead of a policy violation. The errors are designed to be self-explanatory enough for the model to recover.
- Scope by profile. Issue an API key scoped to the profiles the agent should touch, not your whole workspace.
Agent vs webhook: when to use which
Webhook pipelines (auto-reply, AI review responses) are the right shape for high-volume, rules-first automation that runs unattended. MCP agents shine where judgment and context dominate — triaging a mixed inbox, handling the weird 20%, or working interactively next to a human. Most teams end up with both: webhooks for the firehose, an agent for everything the rules don’t cover.
Full tool schemas live at postproxy.dev/automation/mcp; the publishing-side walkthrough is How to Let AI Agents Publish to Social Media via MCP.
