Agentiq

The same workspace the API opens, as tools an agent can use directly. No client library, no glue: point an MCP client at the URL and it can read what the brand knows, teach it something new, see the connected accounts, write posts and read how they did.

The endpoint

https://agentiq.social/api/mcp

Streamable HTTP, with the key as a bearer token:

{
	"mcpServers": {
		"agentiq": {
			"url": "https://agentiq.social/api/mcp",
			"headers": { "Authorization": "Bearer aiq_..." }
		}
	}
}

Some clients still cannot send a header. For those, the key goes in the path:

https://agentiq.social/api/mcp/aiq_...

It is the same door with a worse lock - a key in a URL ends up in logs and in shell history - so use the header where you can, and revoke the key in Settings → API if it ever travels somewhere it should not.

The tools

Tool	What it does
`list_platforms`	The accounts this workspace can post to, and which need reconnecting
`list_posts`	Drafts, scheduled, published and failed, by date and status
`get_post`	One post in full, with its thread and its link
`create_post`	Write a post - a draft, or scheduled
`schedule_post`	Send a draft that already exists
`cancel_post`	Take a post off the queue, or delete a draft
`make_carousel`	Turn a thread into an Instagram carousel - free, seconds
`attach_picture`	Put a picture that already exists onto a post - free
`remove_picture`	Take the attached pictures off again
`draw_post_image`	One branded picture for a LinkedIn or Facebook draft - about four cents
`talk_to_agent`	Talk to the agent that writes the posts, and see what it made
`list_conversations`	The conversations in this workspace
`read_conversation`	One conversation, as words
`read_brain`	What it knows: the brand, the saved facts, the proof, the shelf
`remember`	Save one durable fact about the business
`add_proof`	Add one result this brand can back up
`forget`	Remove a fact or a proof entry by id
`keep_material`	Put a page or a document on the shelf to write from
`read_material`	Read one document from the shelf
`read_results`	What a normal post gets here, and how the recent ones did

Which tool for what

create_post writes exactly the words it is given. talk_to_agent thinks: it has the brand's voice, its proof and its memory, it can search, and it answers with what it SAID and what it MADE, ids included. A client that writes the post itself and pushes it through create_post skips all of that - and it reads like it.

Teaching it things

talk_to_agent already writes from everything in read_brain - the brand's voice, the facts saved about the business, the proof it can stand behind and the calls on its shelf. What the brain door adds is the other direction: an agent that learns something durable can put it back.

Three shelves, and they are not interchangeable:

remember - one plain sentence about the business that stays true. An offer, a lead magnet, a running campaign, a standing preference. Fifty of them, then the oldest have to go.
add_proof - a result with something behind it. This is the shelf every claim is checked against, so a post can say it only if it is here. Mark timeSensitive when the number moves: a follower count is true the day it is written and quietly wrong a quarter later, so those get dated.
keep_material - a page (url) or text you already have (title and text): a transcript, a case study, notes from a call nothing recorded. It lands beside the recorded calls, and the writers reach for it before they invent an example.

The brand itself - colours, typefaces, voice, strategy - is read-only through this door. That one somebody sat down and decided, and a machine quietly rewriting it is how a brand drifts without anybody choosing to change it.

A picture the person already has

MCP tools carry JSON, so nothing can upload a file through this door. That left draw_post_image as the only way to put a picture on a post - four cents, and a different picture than the one they meant.

attach_picture takes the picture's own address and fetches it. Free, immediate, and it beats a drawn picture at publish time. When somebody says they already have the image, that is the tool.

The mistake everybody makes once is pasting the post's address rather than the image's; it answers with what to do about it.

The one rule worth knowing

create_post without a time saves a draft and sends nothing, and schedule_post is the only way one of those leaves later. An agent that decides on its own that the moment has come does not get to publish; somebody has to have asked. Give it a scheduledAt - an ISO time, or "now" - only when they have.

Everything else is the app's own rules, arriving as sentences rather than codes: a network that is not connected, a post with no picture where the network takes no text, a TikTok post whose privacy level nobody chose.