Handler is a social media manager whose primary user is an agent. Connect your accounts once, point Claude at Handler over MCP, and it drafts, schedules and publishes across your networks. Every post it writes waits in a review queue until you approve it.
Publishing on X and Facebook today · listening on TikTok.
It’s handled
Handler runs a remote MCP server, so Claude sees your accounts as tools rather than a website to click through. The tool names below are the ones Handler actually exposes.
you post the launch note to X and Facebook claude list_accounts → 2 profiles create_post → queued for review handler 1 post waiting in /review
The post stops there. It publishes when you approve it, and not before.
The agent rail
Five tools over MCP. Each one maps to a REST endpoint under /api/v1 — the connector and the API are the same spine, not two separate products.
list_accountsEvery connected profile, and what it is allowed to do.create_postOne call, every account you name. Text posts.schedule_postPut it on the calendar. A minute-resolution cron dispatches it.get_post_statusPer-network delivery, including the partial failures.cancel_scheduled_postPull it back before it goes out.The trust layer
The hard part of an agent-first product isn’t the posting. It’s making it reasonable to hand over the keys in the first place.
There is no setting that lets an agent publish unsupervised, because that setting would be the whole risk. Posts arrive as drafts and wait.
Your agent writes; you skim and approve. Approving is the only thing that turns a queued post into a published one.
Platform tokens are encrypted with AES-256-GCM under a per-workspace derived key. Your API keys are stored as hashes — Handler cannot show one to you twice.
Every use of a platform token writes a row: which account, what it did, when. Handing over your accounts should leave a paper trail.
Listening
Handler sweeps social platforms for your keywords every morning, then pulls the comment threads on the posts that matter — which is usually where the actual leads are, not in the post itself.
The hard part is that keyword search on these platforms is not exact match. In one early run, 208 of 214 collected posts didn’t contain the keyword they were filed under — searching “elemental prep” was returning content about a music festival. So Handler verifies every post against its own text before it spends anything analysing it.
AI matches
What survives the keyword check goes to Claude with a description of your brand. It answers three things about every post, and it is told to mark a post irrelevant when in doubt — because the cost of a cluttered inbox is that you stop opening it.
RelevantA yes/no on whether the post is genuinely about you. This is the “AI match” badge.ClassificationReview, question, complaint, competitor, creator lead, or plain mention.SentimentPositive, neutral, negative, or mixed.A comment counts as a match even when it never names you — “is this good for beginners?” under a review of your product is a lead, and it is judged in the context of the post it replies to.
Networks
The honest version. A network appears here when it has a working adapter, not when it appears in the database schema.
Questions
Connect an account, add the connector to Claude, and let it draft something small.