The assistant pattern: a person describes what should go out, an agent does the work of uploading, planning and scheduling, and nothing is created until the person approves. In Dropslate the approval is not a convention the agent is asked to follow; dropslate_create_post refuses without confirm: true, and dropslate_plan_post has no side effects, so the agent can plan as often as it likes and create only what was approved. This recipe is the sequence as it happens in a chat.
A paid plan (on Free the agent can read, not publish) and an MCP client: Claude.ai or ChatGPT over sign-in, Claude Code or Cursor with an API key - Connect an AI agent.
The files: on your computer (Claude Code, Claude Cowork - the agent uploads them from a shell) or at public URLs (any client).
Connected accounts; the agent reads them with dropslate_list_accounts.
and runs the returned upload from your shell (curl -X PUT "<uploadUrl>" -H "Content-Type: video/mp4" --data-binary @monday.mp4), polling dropslate_get_media until ready. Then:
Plan for Friday: (the same table for 2 Oct). Nothing is created yet. Confirm both?
You: Yes, but move Friday to 17:00.
The agent re-plans Friday at 17:00, shows the table, and after your yes calls:
{ "tool": "dropslate_create_post", "arguments": { "...same body as the plan...": "", "confirm": true } }
Agent: Queued. YouTube · What We Made · 28 Sep, 18:00 (Kyiv). Queued. Threads · @whatwemade · 28 Sep, 18:00 (Kyiv). Queued. YouTube · What We Made · 2 Oct, 17:00 (Kyiv). Queued. Threads · @whatwemade · 2 Oct, 17:00 (Kyiv).
The report comes after the tool returned success, with the post id, the account and the absolute time in the workspace’s zone - the format every tool description asks for.
Open Queue: four posts, createdVia: mcp in each drawer. Settings → MCP → Last calls lists every tool the agent used, with the result, and mcp.call rows are in the action log. If the agent had called dropslate_create_post without your confirmation, the tool would have answered confirm_required and created nothing; if it had called it with confirm: true without showing you a plan, the posts would exist - which is why the plan-then-confirm rule is in every tool’s description and why a client that lets you see tool calls is worth using.
Read-only review. On the Free plan, or with a read key, the agent can list accounts, plan and answer “what is scheduled next week” - nothing else; tools/list shows only the 11 read tools.
Correcting a failed post. “Why did Tuesday fail?” - the agent calls dropslate_get_post, reads the network’s message, and proposes dropslate_retry_post (again behind confirm: true) or a corrected re-plan.
A pipeline that asks a human. In n8n, the same shape without an LLM: the plan’s table goes to Slack or email, an approval link continues the flow to POST /posts.