work / queuecat · own product · 2026
QueueCat
Social publishing where the AI is the operator and the human is the editor
system
AI agent (any MCP client)
│ scoped OAuth token — never sees channel credentials
▼
┌─ MCP server — 19 tools ────────────────┐
│ idempotency · scopes · destructive │
│ confirms · rate limits · audit │
└──────────────┬─────────────────────────┘
▼
queue ── guardrails ── APPROVAL (human, phone) ── publish
│
delivery receiptsthe idea
Most 'AI social media tools' bolt a chatbot onto a dashboard. QueueCat inverts it: the agent is the operator. It drafts, schedules, and analyzes through a Model Context Protocol server, and the human approves from a phone. The interesting work is everything that makes that safe.
what makes it production-grade
The MCP server exposes 19 tools, each with idempotency keys so a retried agent call can't double-post, partial-update semantics, and destructive-action confirm handshakes — a tool marked destructive requires an explicit second call to execute.
Auth is OAuth 2.1 with PKCE and dynamic client registration, and API keys are scoped per capability with step-up scopes for risky operations. The agent never holds channel credentials; they stay server-side, behind guardrails the agent cannot talk its way past.
Around the agent: a full REST surface with an OpenAPI spec, webhooks, delivery receipts, cron reapers for stuck jobs, and an approvals engine so a human can accept or reject anything the agent queued.
why it matters if you're hiring
This is the shape of every serious 'let an AI do the work' system: capability through tools, safety through scopes and confirms, trust through human approval. If you want an agent operating inside your business, this is the architecture I'd adapt to it.
stack
Next.js 16 · React 19 · MCP SDK · Better Auth · Drizzle + Postgres · TanStack Query
Private codebase — in active development.