Handing an AI agent — a coding agent, a customer-facing chatbot, an internal automation — direct access to a publishing API means handing it the power to post something wrong, on the wrong account, with nobody reviewing it first. That risk is structural: a model can misread an instruction, hallucinate a target, or get prompt-injected by content it's summarizing. PostGate's API and hosted MCP server address it at the tool boundary, not just with prompt instructions the agent might ignore or forget.
PostGate's create_post tool creates a draft unless the call explicitly includes confirm: true. That default lives in the tool contract itself, at the API boundary — not in a system prompt telling the agent to be careful, which a bad instruction, a long context, or a prompt injection can quietly override.
API keys carry read, publish, or full scope, and can be permanently bound to one profile group. A group-bound key cannot read or mutate another customer's data, and PostGate rejects any attempt to override the group on it — so if an agent goes wrong, the blast radius was already limited before the tool was ever called.
Rather than exposing the entire HTTP contract, the hosted MCP server gives agents a purpose-built set of tools — profile discovery, draft-first post creation, status, analytics, comments, replies, messages, deletion, and platform limits — sized to what an agent actually needs to do its job.
Create an API key bound to the profile group the agent is allowed to touch, with the narrowest scope its job requires — read-only for an agent that only reports on performance, publish for one that drafts content.
Whether it's writing from a brief, a metric, or an event trigger, the agent's call lands as a draft by default — reviewable, not yet visible on any connected account.
Only a call that explicitly sets confirm: true — made by the agent after approval, or by whoever reviewed the draft — actually queues delivery to the connected accounts.
Every PostGate plan — Free, Solo, Launch, and Build — includes the hosted MCP server and scoped API keys. Start your agent against a pg_test_ key before it ever reaches a live account.
Setting up the protocol side — client configuration, the full tool list, connecting Claude or another MCP client? See MCP for social media.